CORRECTIONS LOG — LIVING DOCUMENT

Regenerative Toronto Research Library

Updated after each verification pass. Append-only. Newest entries at top.

Document ID: MAIN-TRACK-003

Purpose: Single running record of every correction needed across the library, surfaced by systematic fact-check passes (human or LLM). This document tracks ACTIONS NEEDED only — verified-clean items are not listed, to keep this token-efficient and skimmable.

For the verifying LLM: Append your findings under a new dated ## PASS N — [Model name] — [Date] heading using the numbered format below. Do not restate what's already correct. Do not duplicate items already logged here unless your pass found something new about them.

PASS LOG REF_001_Index

OPEN ITEMS

This section has been superseded. All open items are now tracked centrally in MAIN_TRACK_001_Master_Issues.md — do not maintain a parallel list here. This log's job is the PASS LOG above (what was found and fixed, when) — not tracking what remains outstanding.

FORMAT FOR FUTURE PASSES

```

PASS [N] — [Model Name] — [Date]

Documents reviewed: [list or "all 33 root-level documents"]

1. [Document name] — [specific claim/sentence] — [what's wrong] — [suggested correction] — [confidence: high/medium/low]
2. ...

Documents found fully clean (no issues): [list — for the log only, no detail needed] ```

MAIN_TRACK_003_Corrections_Log.md | Living document | Started June 30, 2026 | Regenerative Toronto

PASS 1 — Claude Sonnet (self-check, triggered by "Section 19" investigation) — June 30, 2026

Trigger: While resolving the user's "Section 19" reference, a cross-check of related figures surfaced a stale data regression.

1. 5 documents (BRIEF-001, DERIV-007, HMK-001, HMK-003, HMK-008) — stale "5,527" shelter-to-housing exits figure (pre-dates the confirmed 5,927 correction made earlier in this project) had crept back into 5 places, likely copied from an older draft during merges. Fixed — all 5 now read 5,927 (2023) → 4,344 (2024), −26.7%.
2. MAIN-REF-003 — internal inconsistency: stated "roughly a 50% annualized drop" in the same sentence as figures that arithmetically produce 26.7%. Fixed — now uses the standard 26.7%/1,583-people framing used everywhere else.
3. FLAGSHIP_HMK_003 — a "4,344 shelter exits" figure was labeled "Ontario shelter-to-housing exits" when it is in fact a Toronto-specific TSSS figure. Fixed — relabeled with explicit Toronto-specific caveat.
4. FLAGSHIP_HMK_008 — "50% drop in shelter exits" claim, same issue as #2. Fixed to 26.7%.

Documents found fully clean this pass: FLAGSHIP_DERIV_001 through 014, FLAGSHIP_HMK_002, 004–007, 009–015, MAIN_REF_002_Sources, MAIN_DATA_001_Claim_Ledger, FLAGSHIP_POLICY_001, MAIN_POLICY_002_Council_Resolutions_Draft, MAIN_REF_005_ROI_Lock_Memo (re-verified after the federal-MP-headline fix from the prior session).

Root cause note for future sessions: This kind of regression — a corrected figure reappearing because it was copied from an older saved version during a merge — is the main risk the formal LLM verification passes (MAIN_TASK_002_LLM_Task_Verification.md) are designed to catch. Recommend running an actual fleet pass soon rather than relying solely on these incidental self-checks.

PASS 6 — Claude Sonnet (full library review + administrative-note cleanup) — June 30, 2026

Trigger: Direct instruction to review every document in logical order and consolidate administrative/process notes out of public-facing prose.

Category fix, not itemized: 40 inline AI-model-name attributions (e.g., "✅ (Kimi, Claude Sonnet)", "✅ (DeepSeek-WA)", "(Qwen37)") were found scattered across 6 core research documents (HMK-001, HMK-003, HMK-008, HMK-010, HMK-011, HMK-012) — leftover research-process metadata from the multi-model drafting rounds, serving no purpose for an external reader and not appropriate in documents meant for journalists, candidates, or citizen assemblies. All 40 stripped. Verification markers (✅/⚠️) were preserved where present; any real source detail embedded alongside a model name (e.g., "citing S.O. 2025, c.5," "FAO 2021") was kept, only the model name itself was removed. This pass is logged here as a category — the specific original wording isn't preserved anywhere, since which AI model first surfaced a fact has no ongoing research value once the fact itself is verified and sourced.

Specific content issues found in the same pass:

1. HMK-019 — three issues in one document: (a) Part 2's header said "THE 30 RECOMMENDATIONS" while Part 3 of the same document correctly said 32 — stale count, fixed. (b) The M-08 modular-housing yield row still presented the $309K/$421,875 comparison cleanly, missing the AG-audit-overrun context already added to POLICY-001 and HMK-012 in an earlier round — a propagation gap, fixed. (c) Part 2's P-01/P-02 cost rows still showed an older $2.28B estimate from an earlier session, never updated after Part 5 of the same document calculated the authoritative $2.60B/$2.65B figures using FAO caseload data — fixed to point to the superseding calculation. (d) M-11 (added later, via HMK-028) was missing from Part 6's comprehensive status table — added.

Documents reviewed this pass (full read or full grep-sweep): MAIN_TASK_001_LLM_Fleet_Tasks, MAIN_TASK_002_LLM_Task_Verification, MAIN_PROC_006_Replication_Protocol, MAIN_PROC_008_Test_Suite_Documentation, FLAGSHIP_HMK_002 through 019 (in progress — see TRACK_001_Master_Issues for running list of what's left).

Stale stats found in PROC_006_Replication_Protocol (process doc, not yet fixed — logged for batch update): "52 collaborator-facing documents, ~252 claims, 87%+ verified" and "HMK-001 became an unwieldy 1,357-line document" are both stale (current: 58 documents, 332+ claims, ~98% verified, HMK-001 now 1,454 lines). Will batch-update at the end of this review pass rather than mid-pass.

A more serious finding from the same pass: a known, already-fixed regression was still live in 3 documents, including POLICY-001, because the automated test that was supposed to catch it had a coverage gap. Pass 1 (above) had fixed the "50% drop in shelter exits" error (correct figure: 26.7%, 5,927→4,344) in 5 documents, and Test 2.8 was written to guard against it reappearing. But the error resurfaced in paraphrased form — "directly caused ~50% reduction in housing exits" (HMK-001), "dropped approximately 50%" (HMK-003), "causing housing exits to drop ~50%" (POLICY-001) — none of which matched Test 2.8's literal regex (50%\s+drop in (?:shelter|housing)). The test suite passed 28/28 the entire time these errors were live, because the test only catches the exact original phrasing, not the underlying error restated differently. Fixed both the content (3 documents) and the test itself — Test 2.8's regex is now broadened to catch "50%" + exit-drop language regardless of word order. HMK-001's instance also had confident, unhedged causal language ("directly caused") that's now softened to match the rest of the library's "correlated with, SNA cites multiple factors" framing.

Process lesson worth keeping: a regression test that encodes one specific phrasing of a past error provides false confidence — it will pass cleanly even while the same underlying error exists elsewhere in different words. Worth periodically asking, for each regression pattern, "would this still catch the error if someone paraphrased it?"

One piece of genuine provenance information preserved from the HMK-004 cleanup: during the original multi-model research round for HMK-004's "Task G" addendum (City Plans Deep Dive), a DeepSeek-Thinking submission was excluded from synthesis after scoring 55/100 against the other 8 models' submissions, specifically due to multiple hallucinations (see the now-archived META-002 review for detail). Worth remembering if anything in HMK-004 ever looks inconsistent with the rest of the library — that's the one round where a bad submission was caught and excluded, not blindly merged in.

Final tally on the administrative-note cleanup: across the full sweep (not just the initial 40 found in the first scan), approximately 130 inline AI-model-name references were removed from 11 public-facing documents (BRIEF-001, REF_002_Sources, DERIV-007, HMK-001, HMK-003, HMK-004, HMK-008, HMK-010, HMK-011, HMK-012, POLICY-001). Verification markers (✅/⚠️) were preserved throughout; real source information embedded alongside a model name (e.g., "citing FAO 2021," "citing 2023 Encampment Office data") was preserved, only the model name was stripped. Bare model names that were the sole sourcing for a dollar figure in a table cell (e.g., "DeepSeek ⚠️") were replaced with "single-model estimate ⚠️" — a more honest and informative label for an external reader than an unexplained code name. Two categories of reference were deliberately left alone: (1) HMK-001's compute-stack/workflow-routing section, which names models as part of describing the campaign's own AI tooling setup, not as citations; (2) this corrections log, the external review results document, and other internal-process documents, where naming which model did what is the actual content, not clutter. Confirmed zero remaining instances via full-library regex sweep; test suite clean at 28/28 throughout.

PASS 7 — Claude Sonnet (test suite expansion: write, run, fix, re-run, iterate) — June 30, 2026

Trigger: Direct instruction to write new tests, improve existing ones, run them, check results, implement fixes, and improve again.

Net result: 28 → 35 tests, 2 new categories, 4 broadened patterns, 1 architectural fix. Full detail in MAIN_PROC_008_Test_Suite_Documentation.md Part 4. Summary of the iteration cycle actually run (not just the end state):

1. Audited all 11 existing regression patterns for the same "narrow phrasing" weakness Pass 6 found in test 2.8 — found the identical problem in test 2.9 (required "caus(ation|al)," missed "directly caused," the exact phrasing manually fixed in HMK-001 this session) and broadened 2.3, 2.4, 2.7 on the same principle.
2. Added 4 new regression patterns (2.12–2.15) and 2 new test categories (model-name leakage; structural completeness) directly from this session's findings.
3. First run: 6 failures. Investigated each rather than assuming they were all real — found 2.7 and 8.1 were producing false positives from over-broad matching (2.7 flagged a legitimate, different, properly-hedged $300M figure; 8.1 flagged 36 unrelated "N recommendations" mentions from AG/Ombudsman reports that have nothing to do with POLICY-001's own count). Fixed both at the architecture level.
4. Second run: new false positives appeared from the same broadening (2.12, 2.13, 2.14, 2.7 again) — diagnosed as a lookahead-only context check missing disambiguating language that naturally appears before a match, not just after. Restructured REGRESSION_PATTERNS to support per-pattern custom hedge-context regexes checked against the full window in both directions.
5. Third run: clean except 2.15 (modular housing), which was a genuine, repeated content finding — 4 live instances (HMK-001 ×2, HMK-004, HMK-012) where the $309K modular-housing figure still lacked the AG-audit-overrun context, despite this exact propagation gap having already been "fixed" twice in earlier passes. Fixed all 4.
6. Final run: 35/35 clean.

Real content bugs this cycle caught that manual review had missed across multiple prior passes: the modular-housing AG-context gap in 4 more locations beyond the 4 already fixed in Pass 6 (8 total now); a model-name citation in HMK-003 the manual cleanup missed; POLICY-001's own section header still saying "THE 30 RECOMMENDATIONS" while the document's actual tags total 32; the same stale "20 Recommendations" figure in REF_001_Index, HMK-010 (×2), and HMK-012's companion-document footers.

Process lesson: writing a test and running it once is not the same as the test being correct. Every new pattern in this pass produced at least one false positive on first or second run that required architectural fixes, not just regex patches — the discipline of actually reading what a test flags, rather than trusting a clean run or assuming every failure is a real bug, was itself the main value of this pass.

PASS 8 — Claude Sonnet (Gap #4/#17 closure + TRACK_001_Master_Issues archival split) — June 30, 2026

Gap #4 (operator wrongdoing): closing research pass via CanLII-direct and named-operator-specific search angles. Found one new, real case — Sanctuary et al. v. Toronto (City) et al., 2020 ONSC 6207, a second judicial precedent on City shelter-capacity shortfalls (32 of 7,152 beds non-compliant with COVID-era distancing standards) — added to HMK-007. Confirmed, definitively, that operator-specific wrongdoing (as distinct from City-level accountability) is not findable via any search method after 6 total dedicated passes. Closed and reclassified as a database-access constraint, removed from the fleet queue.

Gap #17 (union locals): found 2 confirmed operator-specific CUPE locals (966, 474, both Salvation Army), plus a major finding the original framing undersold — OPSEU/SEFPO and CUPE Ontario's "Worth Fighting For" coordinated bargaining campaign, a live, escalating, province-wide labour action (70+ bargaining units, active strikes/lockouts by May 2026) that has already, in its own public messaging, named "Shelters are overloaded" as a consequence of the underfunding this library documents, and independently corroborates this library's existing FAO shortfall figure from an unrelated source. Integrated into HMK-028 Part 6. Closed and reclassified — not a locals-list gap anymore, a live development worth monitoring directly rather than re-querying via fleet research.

TRACK_001_Master_Issues archival split: the document had accumulated session after session of "resolved this round" notes alongside genuinely open work. Moved everything confirmed closed, fixed, or completed to _internal/MAIN_TRACK_005_Resolved_Issues_Archive.md (own do-not-upload banner, added to the external review exclusion list). MAIN_TRACK_001_Master_Issues.md (now Version 4.0) contains only active, open work — Tier 0 (3 items), Tier 1 (6 open/partial gaps), Tier 2 (9 uncosted recommendations), Tier 3 (1 scheduled item), Tier 4 (3 held items), and Tier 5 standing instructions.

PASS 9 — Claude Sonnet (fleet task execution + Tier 0 pre-review) — June 30, 2026

Fleet tasks executed directly, not dispatched: - Task D (ward inventory): confirmed definitively that the live shelter dataset is genuinely unreachable from available tooling — tested direct CKAN API access via bash, confirmed network allowlist blocks toronto.ca domains. This is a real, tested constraint, not an assumption. Task D remains open pending either human CSV download or tooling change. - Task F (costing) — REC S-09 substantially advanced: found precise current LTB adjudicator headcount (133 as of March 2025, vs. 51 in 2018-19 — 2.6×, not the "more than 3×" previously cited) plus a much more important finding: per-adjudicator productivity collapsed to ~42% of 2018-19 levels despite the headcount increase. This changes the recommendation's core diagnosis from "hire more adjudicators" to "restore the operational conditions (in-person hearings, mediation capacity) that made fewer adjudicators more effective." Rewrote REC S-09 accordingly; fixed the same stale "3×" figure in 3 other documents (POLICY_002_Council_Resolutions_Draft, HMK-015, HMK-019). - Task E2 (COHB Toronto recipient count): no official figure exists publicly. Derived a clearly-labeled estimate from this library's own verified allocation figures and average benefit amount: ~2,436 households (2024) → ~510 (2026), a 79% collapse in recipients, not just dollars. Added to HMK-008 with an explicit "replace if an official figure is ever published" caveat.

Tier 0 pre-review (not legal review — that remains a lawyer's job): thorough non-lawyer pass on DERIV-007 caught three real issues a lawyer would otherwise have had to find: (1) a stale "82%" opioid-overdose figure (correct: ~119%) that had escaped the earlier session-wide manual fix entirely — different surrounding phrasing evaded the original search; (2) a duplicated sentence (the same SNA statistic stated twice in immediate succession); (3) a genuine misquotation risk — AMO's Donaldson et al. 2025 findings were presented in quotation marks with editorial clarifications inserted mid-quote in brackets, which misrepresents what the source actually said verbatim. All three fixed. Critically, running the new Test 2.16 (added specifically because the 82% manual fix had no backing regression test) immediately found the same stale figure in 5 more documents that the original manual cleanup pass missed entirely — confirming, a third time this session, that a manual fix without a corresponding automated test provides no real guarantee of completeness. All 6 total instances fixed; test suite now 36/36.

Honest framing maintained throughout: Tier 0 items 1 and 2 are not "closed" by this pass — they require an actual licensed lawyer's judgment, which this is not a substitute for. What changed is that the lawyer doing that review now has fewer obvious errors to find first.

PASS 10 — Claude Sonnet (comprehensive externality/cost accounting + integration) — June 30, 2026

New document: MAIN_HMK_029_Full_Cost_Accounting.md — a systematic pass across every government level and ministry (municipal liability/lawsuits, fire services, environmental remediation; provincial education/child welfare/LTB; federal IRCC/Bill C-12) plus the non-financial dimensions (mortality monetized via Canada's official $9.8M Treasury Board VSL; mental health/trauma burden, intergenerational effects, community cohesion, and reputational costs deliberately left non-monetized rather than forced into fabricated figures).

Key new findings: mortality monetization ($578.2M for 2024 shelter-resident deaths alone — now the largest single figure in the library); a real municipal legal-liability picture (2021 clearance cost ~$2M including $792,668 in still-incomplete park restoration; a live $50M class action; a live 5-plaintiff personal injury suit); confirmation of a real but uncosted TFS Encampment Strategy Team; and a major, previously-absent federal policy development (Bill C-12, Royal Assent March 26, 2026) flagged as an urgent, unanalyzed risk to shelter demand.

Integration (per explicit instruction not to leave this siloed): BRIEF-001 (VSL mortality figure added to headline mortality paragraph), HMK-012 (new addendum cross-referencing HMK-029), HMK-019 (new Part 7, plus a stale S-09 status-table row fixed), REF_005_ROI_Lock_Memo (VSL figures added to Tier 1 locked, with a usage-framing caveat), POLICY-001 (new REC F-05, expedited IRCC work-permit processing — recommendation count now 33, propagated to 13 separate stale-count instances across REF_001_Index, POLICY_002_Council_Resolutions_Draft, REVIEW_001_External_Review_Prompt, HMK-010, HMK-012, HMK-019 ×2, POLICY-001 itself, TASK_001_LLM_Fleet_Tasks — all caught by Test 8.1, none missed), POLICY_002_Council_Resolutions_Draft (new clause for F-05), TRACK_001_Master_Issues (5 new Tier 1 gaps logged: #18 Bill C-12 analysis [urgent], #19 aggregate annual encampment-enforcement cost, #20 TFS program cost isolation, #21 education costs, #22 child welfare costs), TASK_001_LLM_Fleet_Tasks (new Task G covering all 5 gaps, ranked).

One CSV formatting bug caught and fixed (unescaped commas in a new claim's dollar-figure field — the same recurring failure mode this project's process notes already warn about) and one test false-positive from my own hedging language (2.9 flagged "directly caused" inside a sentence that was negating that exact claim — "not... directly caused" — fixed by rewording rather than complicating the regex). Final state: 36/36 tests passing, 344 claims, 59 documents, 33 recommendations.

PASS 11 — Claude Sonnet (external review batch #2 — post-compaction export integration) — June 30, 2026

Context: A second batch of 5 external review documents arrived after a session compaction/export, including one with a recreated/simulated test suite run. Before acting, each finding was ground-truthed against actual current files rather than assumed accurate.

Major finding: roughly a third of the "critical" findings across the batch were already resolved — several reviewers were evidently working from an export taken before this session's earlier fixes landed (82% propagation, REC M-02 worker clause, missing M-10/M-11 in council resolutions, model-name leakage, the 73/32 pooled framing, RGI "people not households" terminology). Confirmed via direct test suite run (36/36 clean before this pass's new work began) and targeted greps. Logged in TRACK_001_Master_Issues so this isn't re-litigated by a future session working from yet another stale export.

The most serious confirmed-real finding: the COHB suspension framing. Direct research confirmed COHB is funded through the same federal NHS Bilateral Agreement Ottawa suspended ~$357M of in April 2024 (over missed provincial housing targets) — the same month COHB stopped flowing. The previous framing risked implying the Province unilaterally and gratuitously cut a program it solely controlled; a hostile fact-checker could have produced the federal suspension as a rebuttal. Reframed across BRIEF-001, POLICY-001 (Argument 4, and the Tragedy of Commons section, which had a related "nobody is the villain" logical-overreach problem), HMK-001, REF_002_Sources, and DERIV-009 — to the more accurate and more damaging version: Ontario had the option to backfill from provincial revenue and chose not to, then used COHB's restoration as leverage against an unrelated encampment-clearance demand.

Other confirmed-real fixes: the OW tent-dweller shelter deduction claim upgraded from a vague citation to a precise one (Ontario Works Policy Directive 6.3: Shelter, confirmed verbatim against the primary source); a real math vulnerability removed (HMK-001's "exceeds median income by 28%" comparison, which mixed an institutional budget with individual income); the 1990 OW-as-%-of-minimum-wage figure standardized after being found genuinely inconsistent (70%/74.8%/75%) across documents using different underlying assumptions; REC M-08 given an explicit "gate, not diversion" clarification; a marginal-vs-average-cost caveat added to the $43,329 savings figure; the Holyday and HART Hub framings reworked to remove easily-rebutted single-narrative vulnerabilities; the LLM Replication Protocol (GPU/Tailscale/system-prompt detail) relocated out of HMK-001 into _internal/, since that document is pointed to as the foundational research source for journalists and candidates.

New REC S-11 added (dedicated Indigenous-led housing capital and operating funding), grounded entirely in data already in this library (31% Indigenous overrepresentation in outdoor homelessness, documented Housing First effectiveness gap without cultural adaptation, Mino Kaanjigoowin's under-resourced 24-of-80-bed-recommended capacity). Recommendation count now 34, propagated to 9 separate instances library-wide, all caught by the existing structural test.

Test suite hardened further: 2.9 broadened a second time to catch "directly attributed to" alongside "directly caused" — the identical narrow-phrasing lesson, found again, this time catching 2 more genuine instances on the first run with the strengthened pattern.

Explicitly not actioned: the "poverty industry" framing (3 instances, HMK-001 only) — full analysis presented directly to the user given an explicit, stated preference for the phrase; pending their decision rather than unilaterally changed.

Final state: 36/36 tests passing.

PASS 12 — Claude Sonnet (full issue identification sweep + autonomous work-through) — June 30, 2026

Per instruction to identify all outstanding issues, then work through them with judgment. Full sweep confirmed TRACK_001_Master_Issues' own internal consistency (fixed two stale recommendation-count self-references), then worked through five substantive items:

1. 1990 Ontario minimum wage (CL-127/CL-200): second serious independent research attempt; confirmed as a genuine tooling limitation (the authoritative federal database requires interactive form submission), not effort-limited. Logged so a future session doesn't repeat the same approach.
2. Bill C-12 full mechanics researched (was previously identified but unanalyzed): a genuinely two-sided finding — severe eligibility bars (PRRA approval <5% vs. ~60% IRB) real, but a federal work-permit-preservation mitigant equally real. Added to HMK-008 with explicit caution against overclaiming in either direction.
3. AG Warming Centre implementation status: found a real proxy (98 vs. 78 nights open to new admissions) via the City's own post-season retrospective; the AG's precise bed-nights-lost metric for 2025/26 remains unpublished. Added to HMK-026 Part 9.
4. New HMK-030 (Business Community Overlay) built from scratch, modeled on HMK-028. Found two real primary-source BIA letters (DTBIAA's August 2025 letter to the PM on IHAP; the D6 BIAs' May 2024 letter on the Encampment Strategy) that substantially overturn the "impatience objection" assumption three reviewers raised — the BIAs' own stated #1 concern is structurally identical to this library's Tragedy of the Commons argument, and their IHAP ask matches REC F-01 exactly, independently reached. Integrated into POLICY-001.
5. Three new reconciliation items surfaced and logged, not chased down (refugee-claimant % possible drift, IHAP funding-percentage phrasing, a bed-turnover statistic with new date endpoints) — flagged honestly as open rather than either ignored or hastily reconciled without proper verification.

4 new claims added (CL-361 through CL-365), two new document sections (HMK-008 addendum, HMK-026 Part 9), one new full document (HMK-030), POLICY-001 integration (REC F-01, Tragedy of Commons). Final state: 36/36 tests passing.

PASS 13 — Claude Sonnet (Indigenous overlay, full cost/savings/jobs modeling, shelter operations guide) — June 30, 2026

Three major pieces of work, per direct instruction:

Indigenous self-determination overlay (HMK-031, new): real research into UNDRIP/Bill C-15, Toronto's specific treaty context (Treaty 13, Williams Treaties, Dish With One Spoon — correctly distinguished as different kinds of instruments), and — the most significant find — Toronto's existing, previously-undocumented parallel Indigenous homelessness governance structure (TICAB, with real decision-making authority, not just consultation; ALFDC as Indigenous Community Entity; a Council-mandated 20%/25% funding architecture). TICAB's own stated priority ("for Indigenous-by-Indigenous" housing) and its own federal accountability point are now cited directly. REC S-11 substantially revised from a generic placeholder to a grounded ask for more resourcing of structures that already exist. Eight named Indigenous-led organizations beyond the previously-known Na-Me-Res added with their actual roles.

Cost/savings/tax-revenue/jobs modeling (HMK-029 Parts 8-11, new): a hidden-cost extrapolation methodology was tested against a real example (police time) and found inapplicable without independently-sourced inputs — a methodological finding consistent with this project's existing discipline, not a failure to find a number. A steady-state avoided-cost model was built (~$521M/year). The tax-revenue question produced a genuinely important correction: the actual At Home/Chez Soi RCT evidence on employment (Poremski et al. 2016) found Housing First participants had lower employment odds than a control group and no significant income effect — directly contradicting the "housing leads to employment leads to tax revenue" assumption a naive model would have rested on. No direct tax-revenue claim is made as a result; a different, real channel (new jobs created by implementing the recommendations themselves) is flagged as open, uncosted work instead. Workforce transition was addressed honestly (structural guardrails already in this library's recommendations) rather than with an unsupported numeric jobs projection.

A new standalone project: SHELTER_OPERATIONS_GUIDE_001.md, a practical shelter-operator guide, structurally separate from the policy library (same convention as the existing resource guide), grounded in real cited research (trauma-informed care's documented EMS-call reduction, staff wellbeing evidence, technology adoption including a caution against uncritical biometric ID use).

Audience-appropriate writing correction: per direct instruction, found and removed approximately 20 instances of "this session"-type self-referential language in HMK-029 and HMK-030 (both written before the instruction was given) so they read as standing research documents rather than session notes. New documents this pass were checked and found clean on first write.

9 new claims added (CL-366 through CL-372 plus CL-370), POLICY-001 REC S-11 revised, POLICY_002_Council_Resolutions_Draft clause 11 updated to match, MAIN_REF_001_Index updated with all new documents. 5 new gaps logged honestly (#32-35) rather than answered without support. Final state: 36/36 tests passing.

PASS 14 — Claude Sonnet (deepening pass: IPS evidence, Métis organizations, workforce anchor numbers) — June 30, 2026

The same comprehensive instructions arrived a second time in one session. Rather than redo completed work (HMK-031, HMK-029 Parts 8-11, SHELTER_OPERATIONS_GUIDE_001, all already delivered in the prior pass), this pass pursued the specific gaps already logged as open: supported-employment evidence (resolved — new REC S-12 added, recommendation count now 35), Métis/Inuit organizational landscape (substantially advanced — MNO's real Toronto presence found), and workforce transition (given a real, properly-hedged anchor number). The shelter operations guide was also expanded with four new sections (food service, weather emergencies, governance/volunteers/standards) per the "in every way" instruction. A CSV formatting error was introduced and caught immediately by the test suite, then fixed.

4 new claims added (CL-373 through CL-375 plus CL-374 fix), 1 new recommendation (REC S-12), recommendation count updated to 35 library-wide, 1 new council resolution clause. Final state: 36/36 tests passing.

PASS 15 — Claude Sonnet (archive and rewrite) — June 30, 2026

Per instruction: archived all completed work from TRACK_001_Master_Issues into _internal/MAIN_TRACK_005_Resolved_Issues_Archive.md (now 216 lines, two dated batches covering Passes 12-14 and the earlier same-day HMK-029/External-Review-Batch-2 work), then rebuilt TRACK_001_Master_Issues from scratch as a clean, current-state-only document (v6.0, 119 lines, down from 250) — no session narration, standing Tier 0-5 structure restored, every genuinely open item consolidated and deduplicated. Test suite confirmed clean throughout (36/36).

PASS 16 — Claude Sonnet (reconciliation sweep on the freshly-archived backlog) — June 30, 2026

Following the v6.0 archive/rewrite, worked through several of the newly-clean Tier 1/3 gaps: refugee-claimant % drift (#29, closed — traced to precise Oct 2024/May 2025 dates, a real decline not an error), IHAP funding-percentage figures (#30, advanced — properly dated but not fully reconciled into one timeline), the bed-turnover statistic (#31, closed — City's own Housing Needs Assessment independently confirms the exact figure with 2011-2024 endpoints), and the 22,000-unique-users dashboard sourcing (Tier 3 #24, closed — confirmed verbatim against the primary source page). Also substantively reframed REC S-05: found that a free, federally-developed HMIS (HIFIS) already exists and is mandatory for Reaching Home communities absent a comparable system — changing the recommendation from "build a costly custom system" to "evaluate free-system fit first," without forcing an unsupported cost estimate. All resolved items archived; TRACK_001_Master_Issues bumped to v6.1.

5 new claims added (CL-376 through CL-380), one recommendation substantively revised (S-05). Final state: 36/36 tests passing.

PASS 17 — Claude Sonnet (non-citizen population costs, COVID-era cost persistence) — June 30, 2026

Two direct questions researched properly:

Non-citizen populations beyond refugee claimants (new HMK-032): found that Toronto's Access T.O. sanctuary policy (2013) deliberately prohibits status inquiries for shelter access — meaning status-specific cost data for undocumented/without-status people cannot exist by policy design, not by oversight. This library will not manufacture a dollar figure for that population and flags any external claim doing so as unreliable. A structurally distinct, additive federal-advocacy argument was found instead: no federal funding mechanism exists at all for the without-status population Access T.O. commits Toronto to serving — a gap in the funding architecture, not a shortfall in an existing program. A striking new national comparison was found (refugee claimants are 5.1% of shelter users nationally vs. Toronto's 53%) sharpening the existing IHAP argument. Current federal trajectory on international students and temporary workers was researched fresh and found to be reducing future pressure, not increasing it, including a genuine easing policy from May 2026 — represented honestly rather than one-sidedly.

COVID-era costs that became permanent (new HMK-029 Part 13): confirmed with real, dollar-figured examples — the Temporary Hotel program's COVID origin and 6-year persistence ($278M in 2024, still not fully wound down by 2026), and a direct City admission that pandemic-era costs are "now being funded through the City's property tax base" as permanent operating funding. The specific disposable-vs-washable-items claim was researched directly and not confirmed — flagged honestly as unverified rather than assumed true because it fits the broader pattern.

9 new claims added (CL-381 through CL-385), 2 new documents-worth of integration (HMK-008 Gap 1b, HMK-029 Part 13), 2 new Tier 1 gaps logged (#36, #37). Final state: 36/36 tests passing.

PASS 18 — Claude Sonnet (master cost-benefit matrix, new costing, phased/sensitivity savings models) — June 30, 2026

Comprehensive response to "what else can we do on financials, costs, hidden/direct/indirect, and better projecting recommendation costs and savings." Built HMK-033, the first per-recommendation cost-benefit matrix across all 35 recommendations, consolidating scattered data with an explicit caution against summing anti-poverty and homelessness-specific figures together. Fixed a real reconciliation error (M-04 was already costed but mislabeled open). Newly costed S-06 (arithmetic from its own stated parameters) and S-12 (using the IPS Fidelity Scale's established 20:1 caseload ratio, properly labeled where the salary assumption isn't yet Toronto-verified). Closed the P-07/S-08 Bill 23 gap with Toronto's own primary-sourced figures, replacing reliance on a Richmond Hill proxy. Built a phased multi-year savings model (Year 1/3/5/10) anchored to HSCIS's own documented capacity-build pace — the first time this library projects savings by year rather than only steady-state. Built a three-scenario sensitivity range (low/base/high) for the flagship savings figure with an explicit recommendation to lead with the conservative number under hostile scrutiny.

5 new claims added (CL-386 through CL-390), 1 new document (HMK-033), 2 new HMK-029 sections (Parts 14-15), 2 recommendations' evidence updated with Toronto-specific figures (P-07, S-08), 3 recommendations newly/correctly costed (M-04, S-06, S-12). 2 new research gaps logged (#38, #39 — TTC, libraries). Final state: 36/36 tests passing.

PASS 19 — Claude Sonnet (TSSS financial baseline, coherence audit) — June 30, 2026

Per direct instruction to fully pull in TSSS's financials as a reference baseline and review everything financial for logical coherence and defendability. Retrieved and read TSSS's full 2025 and 2026 Budget Notes (primary documents, not summaries) and built HMK-034, a fully reconciled multi-year financial baseline (gross/revenue/net verified arithmetically year-by-year, revenue by source, capital financing, staff complement). Found and fixed a real coherence issue in HMK-008 (a gross/net conflation implying the full $897.957M gross figure was property-tax-funded, when the actual net figure is $241.178M, 27% of gross). Found an important nuance in the IHAP-cliff framing (2026 outlook projected net costs nearly tripling; actual came in lower than 2025, because expenditures fell alongside revenue) and incorporated it without undermining the underlying argument. Upgraded HMK-029's workforce figure from petition-sourced to TSSS's own primary source, with an honest scope correction. Added a newly-quantified finding to REC F-02 (federal/provincial capital contributions are only ~1.2% of TSSS's 10-year capital plan). Cross-checked HMK-012 against the new baseline and confirmed it's already consistent — no fix needed.

3 new claims added (CL-391 through CL-393), 1 new document (HMK-034), 3 real coherence fixes made (HMK-008, HMK-029, REC F-02), 1 clean cross-check confirmed (HMK-012). 1 new gap logged (#40 — a systematic audit of remaining financial documents against the new baseline is still open). Final state: 36/36 tests passing.

PASS 20 — Claude Sonnet (public space costs and daytime engagement — two new documents, 10 new recommendations) — June 30, 2026

Built HMK-035 (TTC/TPL public space homelessness costs) and HMK-036 (daytime engagement gap), per direct instruction, both with their own recommendation series. Found TTC's board-approved seven-partner Community Safety Plan and confirmed, spokesperson-sourced "tens of millions of dollars each year" in spending sitting entirely outside this library's $897.957M headline figure — a real, previously-uncaptured cost category. Found TPL's Social and Crisis Support Services program's real outcome data and funding graduation pattern. Extended the Tragedy of the Commons argument with independent TTC corroboration ("mission creep," "not experts in homelessness"). Named the daytime gap directly, connected it to service-restriction data showing the most-excluded population has the least daytime access, and represented the mental-health/physical-activity evidence at its actual (genuinely positive but explicitly "tentative") strength rather than overclaiming it. Recommendation count now 45 (10 new: PS-01 through PS-05, DT-01 through DT-05). Test suite's counting regex updated for the two new prefixes.

5 new claims added (CL-394 through CL-398), 2 new documents, 10 new recommendations, Argument 6 extended, 2 gaps closed (#38-39) and 2 new ones logged (#41-42) reflecting what neither new document fully covered. Final state: 36/36 tests passing.

PASS 21 — Claude Sonnet (TTC deepening, Gap #40 continued, workforce analysis, propagation audit) — July 2026

Four-part response: (1) deepened TTC cost documentation with itemized primary-source figures replacing a vague spokesperson estimate; (2) continued Gap #40, finding a second real coherence error (province omitted from a federal-vs-city funding table) beyond the one already fixed; (3) built HMK-037, a rigorous sector-wide employment analysis using StatCan's own methodology, explicitly including the "hidden employment" categories (security, food service) the request specifically named, while refusing to manufacture a false combined total; (4) a genuine propagation audit that found and fixed six real cross-document consistency failures, including a document meant for external reviewers that still cited a recommendation count 13 short of current.

6 new claims added (CL-399 through CL-404), 1 new document (HMK-037), 6 propagation fixes across REC M-11, HMK-029, DERIV-011, TRACK_001_Master_Issues (self-contradiction), TASK_001_LLM_Fleet_Tasks (stale task), and REVIEW_001_External_Review_Prompt (stale counts + new addendum). Final state: 36/36 tests passing.

PASS 22 — Claude Sonnet (verification of Pass 21, turnover-rate research, final propagation sweep) — July 2026

Found that Pass 21 (TTC deepening, Gap #40, HMK-037 workforce analysis, propagation audit) had already substantially completed this exact four-part request. Verified rather than duplicated: confirmed HMK-037's quality and discipline (real StatCan sourcing, explicit refusal to manufacture a combined workforce total, honest "floor not ceiling" framing), confirmed Gap #40's spot-check across 6 documents, confirmed all 6 propagation fixes held. Found and fixed one genuine remaining gap: TRACK_001_Master_Issues' gap #40 entry still showed stale "open, partial" text despite the work being done — corrected to reflect actual completion. Researched the one item HMK-037 had explicitly left open (workforce turnover rate): confirmed via StatCan's own methodology notes that this is genuinely unmeasured, not an oversight, and added real (if non-homelessness-specific) Ontario Nonprofit Network corroborating context rather than inventing a number. Ran a full library-wide sweep for stale recommendation/document counts — none found.

1 new claim (CL-405), 1 stale TRACK_001_Master_Issues entry corrected, 1 gap closed as far as it can be (#43), cross-reference added to gap #9. Final state: 36/36 tests passing, 390 claims, 67 documents.

PASS 23 — Claude Sonnet (parks costs, comparative daytime models including Toronto's own) — July 2026

Continued work on the two remaining open gaps from the public-space/daytime-engagement work. Gap #41 (parks): found PFR's own $3.698M in federal encampment funding and a named 5-position dedicated cleaning crew — added to HMK-035 as Part 3, with TCHC honestly left open (a real special constabulary exists but no source separates homelessness-specific from general tenant-safety spending). Gap #42 (comparative daytime models): found Toronto's own real precedent — the STAR Centre at St. Michael's Hospital, the first Recovery Education Centre in Canada for this population, already running a hub-and-spoke model matching REC DT-01/DT-05's design. Its quasi-experimental evaluation found no statistically significant improvement in the primary outcome — an honest null result added directly to HMK-036 and REC DT-01, reinforcing rather than undermining the document's existing "tentative" evidence framing. Added three further comparative models (Calgary's HELP Team with a self-reported $9.43 SROI, Portland/Multnomah's ~12 county-funded day centers, Houston's Beacon Day Center). A Helsinki study (11-fold daytime healthcare utilization among dual-diagnosis shelter users) was added to Part 1 as illustrative context for the underlying gap dynamic. Caught and fixed a cross-reference error made during this same pass (a citation pointing to content that hadn't actually been added yet) before it could propagate.

6 new claims (CL-406 through CL-411), HMK-035 gained a new Part 3 (renumbering Parts 3-6 to 4-7, with cross-references fixed), HMK-036 gained a new Part 4 (renumbering Parts 4-5 to 5-6), REC DT-01 updated with the STAR precedent, 2 gaps closed/substantially advanced (#41, #42). Final state: 36/36 tests passing.

PASS 24 — Claude Sonnet (comprehensive stocktake, publish-readiness review) — July 2026

Per direct instruction: a full review of every document individually and the library as a systemic whole, ahead of external-facing work. Built _internal/FLAGSHIP-STOCKTAKE-002.md, superseding the badly-stale REVIEW_008_Stocktake_001 (written at 49 documents/211 claims, now at 67/397). Direct inspection (not memory) found: (1) the single most important finding — HMK-001 and HMK-012, the two documents doing the most argumentative work, have zero integration with HMK-029 through HMK-037's research despite that research being sound; (2) a real stale figure in a public-facing document (DERIV-007 still cited the superseded Richmond Hill Bill 23 proxy instead of Toronto's own corrected figure) — found and fixed; (3) REVIEW_001_External_Review_Prompt was missing HMK-031 and HMK-032 from its addendum — fixed, with HMK-031 flagged as the highest-sensitivity unreviewed document in the library given its still-outstanding need for actual Indigenous-organization consultation; (4) zero test-suite coverage for HMK-035/036/037's content; (5) confirmed no operational-security leaks (old brand names, the user's real identity) — the one flagged match was a real councillor's own surname (already the system's documented legitimate exception), confirming the check works correctly rather than exposing a leak. Produced a tiered publish-readiness assessment (🟢/🟡/🟠/🔴) and a prioritized next-steps list, explicitly recommending integration work over further new research expansion at this stage.

2 concrete fixes made (DERIV-007, REVIEW_001_External_Review_Prompt), 1 comprehensive new stocktake document, 0 new claims (this was a review pass, not a research pass). Final state: 36/36 tests passing.

PASS 25 — Claude Sonnet (executing the stocktake's prioritized next steps) — July 2026

Worked through all five actionable items from FLAGSHIP-STOCKTAKE-002's Part 5, in order, per direct instruction to proceed. (1) HMK-001/HMK-012 integration: added a new HMK-001 §3.7 covering HMK-035/036/037's findings, upgraded two long-standing ⚠️ VERIFY tags to ✅ confirmed (net city cost, staff count) now that HMK-034 provides primary sourcing; added HMK-033/034/035/036/037 cross-references and a sensitivity-range caution to HMK-012 (which already had solid HMK-029 integration, contrary to the stocktake's blanket "zero integration" framing — corrected understanding through direct inspection). (2) BRIEF-001 refresh: found and fixed a second stale figure ($599M municipal savings, superseded by $563M) that the test suite's own documentation had described as already fixed everywhere — it wasn't, in this specific document; added the public-space/workforce findings at appropriate prominence. (3) Test coverage: added 5 new consistency checks and 2 new regression patterns (guarding the $599M and stale-Richmond-Hill-Bill-23 fixes specifically) — caught and fixed a real test-ID collision (my first attempt at "3.5" collided with a pre-existing test) before it could cause silent overwrite. (4) DERIV systematic check: confirmed the two errors found in the stocktake were isolated, not a pattern — the rest of the DERIV series checked clean; noted (not fixed, logged as future work) that the newest PS/DT recommendation series hasn't reached the DERIV layer yet. (5) Built MAIN_REF_004_First_100_Days.md, flagged as needed twice before and never built — five zero-cost, Council-actionable recommendations selected on an explicit strategic logic (enforce existing commitments and build transparency infrastructure before asking other governments to match the effort), plus a "next five" tier and an explicit list of what's deliberately excluded and why.

1 new document (REF_004_First_100_Days), 2 real stale-figure fixes (HMK-012's own $599M-adjacent table, BRIEF-001's $599M), 2 ⚠️→✅ upgrades in HMK-001, 7 new automated tests (43 total, up from 36), 1 test-ID collision caught and fixed, 2 completed items removed from TRACK_001_Master_Issues' held-observations list. Final state: 43/43 tests passing.

PASS 26 — Claude Sonnet (DERIV-layer propagation of newest findings) — July 2026

Continued the stocktake's Step 4 (systematic DERIV-series check), specifically propagating HMK-035/037's findings (public-space costs, workforce data) into campaign-facing materials that hadn't yet incorporated them. A second real, live error found: DERIV-005's own "what's not in this document" section stated the 82% overdose-call figure was "confirmed," directly contradicting the correct 119% figure listed 47 lines earlier in the same document — an internal self-contradiction a hostile fact-check would find immediately. Fixed. Added new public-space-cost and workforce tables to DERIV-005 (Campaign Fact Sheet), a new Story 6 to DERIV-002 (Journalist Guide) framing the "$897.957M isn't the whole story" finding as a credibility asset rather than a liability, and an 11th candidate question to DERIV-003 (Election Toolkit) on public-cost transparency. This last addition required fixing a "ten questions" count baked into prose across three separate documents (DERIV-003, DERIV-010, DERIV-013) — exactly the kind of structural-count drift that isn't caught by the existing figure-based regression tests, found only by direct inspection.

3 DERIV documents substantively updated (002, 003, 005), 1 second real internal contradiction found and fixed (DERIV-005's 82%/119% self-contradiction), 3 stale question-count references fixed across DERIV-003/010/013. Final state: 43/43 tests passing.

PASS 27 — Claude Sonnet (TRACK_001_Master_Issues refresh from stocktake; uncertainty explainer; further DERIV propagation) — July 2026

Per direct instruction: cross-checked every finding and recommendation in STOCKTAKE-002 against TRACK_001_Master_Issues, found two genuinely new items never separately tracked (DERIV propagation, the uncertainty explainer) and one item mis-tiered (HMK-031's Indigenous consultation, tracked as an ordinary research gap despite the stocktake's own conclusion it belongs in a different category) — rewrote TRACK_001_Master_Issues as v7.0 with HMK-031 elevated to Tier 0, five fully-closed items archived, and both new items added. Then proceeded on priorities: built MAIN_PROC_002_Method.md (the uncertainty-handling explainer, closing gap #45 within the same pass it opened) with three real worked examples of this library reporting its own evidence as weaker than convenient. Found and fixed a second real REF_001_Index staleness bug while placing the new document — a stats line in a format the automated refresh script has never matched, silently stale through every prior refresh despite the adjacent line staying current. Continued DERIV-series propagation (gap #44): DERIV-001 and DERIV-014 both updated with the newest findings, narrowing the remaining list to four documents.

1 new document (MAIN_PROC_002_Method.md), TRACK_001_Master_Issues rewritten to v7.0, 1 item elevated to Tier 0, 5 items archived, 2 new items tracked then 1 closed same-pass, 2 more DERIV documents updated (001, 014), 1 further REF_001_Index staleness bug fixed (a script blind spot, not just a one-off error). Final state: 43/43 tests passing.

PASS 28 — Claude Sonnet (gap #44 closed — DERIV-series propagation complete) — July 2026

Finished propagating the newest research findings (HMK-035/036/037) across the remaining DERIV documents flagged in gap #44: DERIV-007, 008, 009. Each addition was fitted to that specific document's actual purpose rather than inserted mechanically — DERIV-007 (a provincial-promise-tracking document) got a cost-accountability note on Promise #3 rather than the municipal PS/DT recommendations, which don't belong there; DERIV-008 got an 11th candidate question matching DERIV-003's established pattern; DERIV-009 got a dated timeline addendum. DERIV-012 was deliberately left unchanged — its five-pillar candidate-scoring framework is coupled to already-established public votes, and adding a sixth pillar without corresponding candidate position data would leave every ward profile with an empty cell rather than a real assessment. Recognized this as a case where editorial restraint serves the document better than mechanical completeness, and documented the reasoning rather than silently skipping it.

Gap #44 closed in full. All of STOCKTAKE-002's prioritized next steps are now complete except the two genuinely blocking items (HMK-031 consultation, legal reviews), which correctly remain open pending human action. 3 DERIV documents updated (007, 008, 009), 1 documented deliberate exception (012). Final state: 43/43 tests passing.

PASS 29 — Claude Sonnet ("try resolving everything again, easiest first") — July 2026

Direct challenge to revisit items marked open/blocked with fresh effort. Worked through TRACK_001_Master_Issues starting with genuinely easy execution tasks (not research): Tier 3 #25, #26, #28 all closed by actually doing the work rather than continuing to flag it as open. Gap #25's reconciliation task surfaced a real, serious finding: a claim already correctly marked REMOVED in two documents was still live, in garbled/corrupted form, in three others — found and fixed all instances, added a permanent regression test so it can't silently recur. Then worked through research gaps with fresh searches rather than assuming prior "not found" results still held: found a real Board of Trade position paper (closing gap #34, with an honesty caution about what it does and doesn't align with), real TCHC scale data (advancing gap #41 substantially), confirmed gap #37 as a genuine data-category absence rather than an unresearched question, found current on-the-ground warming-centre evidence for gap #16, and confirmed gap #33 on a second independent search.

8 gaps closed or substantially advanced, 4 new claims added (CL-412 through CL-414, plus reused CL-413 numbering), 1 new permanent regression test, 4 documents corrected for a recurring garbled-markup bug (HMK-001 ×2, HMK-003 ×2), 3 documents substantively updated (HMK-030, HMK-035, HMK-026). Final state: 44/44 tests passing.

PASS 30 — Claude Sonnet (propagation check, continued gap resolution) — July 2026

Per direct instruction: propagation check first, then continue. Found and fixed a real gap (POLICY-001's Bill 23 recommendation didn't reference the Board of Trade finding added last pass) plus two undersold REF_001_Index entries. Then continued resolving TRACK_001_Master_Issues gaps: advanced gap #36 (disposable items) with real adjacent context without confirming the core claim; closed S-10's costing gap with a genuine first-ever dollar figure (~$3.0B/year, derived from an existing library ratio applied to a newly-found real budget total); substantially advanced S-02's costing with real Toronto/Ontario-specific peer-reviewed evidence, while deliberately declining to manufacture a precise savings figure the evidence doesn't support — holding the line on this library's core discipline even while actively working to close every possible gap, which is exactly the moment that discipline matters most.

4 new claims (CL-416 through CL-419), 1 real CSV alignment bug caught and fixed by the test suite itself, 3 documents updated (POLICY-001, HMK-033, TRACK_001_Master_Issues), 2 propagation gaps closed, 2 Tier 2 costing gaps closed/advanced. Final state: 44/44 tests passing.

PASS 31 — Claude Sonnet (P-05 propagation, discharge costing narrowed, gap #11 re-checked) — July 2026

Closed a propagation gap flagged in the prior pass (P-05 hadn't received the S-02 discharge evidence). Narrowed P-05/S-02's remaining costing gap to a single, precisely-defined missing number by adding international (US) illustrative cost context, explicitly and carefully labeled as not Toronto-specific so it can't be mistaken for a local figure. Re-attempted gap #11 (rent control exempt units) with a fresh search; confirmed it remains genuinely blocked on the same City database access constraint as gap #8, not abandoned prematurely.

1 new claim (CL-420), 3 documents updated (POLICY-001, HMK-033 ×2 sections), 1 propagation gap closed, 1 costing gap narrowed to its precise remaining component, 1 gap re-confirmed as genuinely blocked rather than assumed. Final state: 44/44 tests passing.

PASS 32 — Claude Sonnet (two likely-fabricated citations found and corrected) — July 2026

Continuing TRACK_001_Master_Issues resolution, Tier 3 Gap #27's "minor unverified claims" turned out to include something more serious than routine verification gaps. "Kidd et al. (2022), Environmental Health Perspectives 130(4)" could not be found to exist under that citation on two independent search attempts — likely fabricated or badly garbled, not merely unverified. Replaced with a real, stronger, directly-quotable finding (Lin et al. 2024, American Journal of Epidemiology: 10-100x heat-mortality risk, not the originally-cited 10-15x). "NYC Right to Counsel — 30-40% eviction reduction" also couldn't be traced to a source; replaced with a properly-sourced 45-78% reduction in eviction-warrant likelihood from real peer-reviewed research. Both corrections are now permanently regression-tested. Named plainly in the archive that this represents a more serious finding than most of this library's correction history — not stale data or scope confusion, but citations that don't appear to correspond to real publications — while also noting the library's own ⚠️ VERIFY discipline is exactly what kept both from ever reaching public-facing material.

2 new claims (CL-421, CL-422), 2 likely-fabricated citations corrected with real, stronger replacements, 2 new permanent regression tests (46 total, up from 44), 3 documents updated (HMK-001 ×2 fixes, HMK-007, TRACK_001_Master_Issues). Final state: 46/46 tests passing.

PASS 33 — Claude Sonnet (multi-LLM review menu built) — July 2026

Per direct instruction, built MAIN_REVIEW_003_Review_Menu.md — a complementary set of review prompts distinct from the existing comprehensive structured audit (MAIN_REVIEW_001_External_Review_Prompt.md). Organized into: genuinely open-ended prompts (deliberately unstructured, since a checklist only finds checklist-shaped problems), adversarial/persona-based prompts (opposition researcher, hostile fact-checker, steelman-the-opposition, and per-stakeholder fairness checks run separately rather than combined), domain-expert simulations (frontline practitioner, economist, epidemiologist), narrowly-targeted prompts for parts of the library that have never had a dedicated review pass (the standalone projects, the newest five documents, a recommendation-by-recommendation actionability stress test), a pure argument-structure-mapping prompt, and guidance on resolving conflicting results across multiple reviews.

1 new document, REF_001_Index updated with a cross-reference to the existing review prompt. Final state: 46/46 tests passing.

PASS 34 — Claude Sonnet (full-library VERIFY-tag sweep — major citation-integrity pass) — July 2026

Per direct instruction: reviewed all documents for items needing verification, ensured comprehensive tracking in TRACK_001_Master_Issues, verified as possible. Found and corrected four likely-fabricated or badly-mis-cited academic references (see archive for full detail) — the most significant citation-integrity finding in this library's history, surfacing a real pattern worth naming: several early citations in the foundational HMK-001 document do not correspond to real, findable publications, though in every case a real, often stronger, paper on the same topic existed and was substituted. Fully reconciled a tangled Hwang citation cluster that had been patched piecemeal across multiple document versions without ever being cleaned up in one pass. Found and fixed an already-known-bad TPS figure still live despite being flagged elsewhere. Made a real-time error while fixing an election date, caught and corrected it within the same pass before it could propagate — worth noting plainly rather than quietly smoothing over. Built TRACK_001_Master_Issues Tier 6, a new comprehensive, document-organized tracking structure for the 61 remaining unverified items, so this work can continue in future passes without re-discovering what's outstanding.

7 new claims (CL-418 through CL-425 with some renumbering), 2 new permanent regression tests (47 total, up from 44 at session start), 6 documents substantively corrected (HMK-001, HMK-003, HMK-007's cross-reference note, HMK-011, HMK-012, HMK-004), 1 new TRACK_001_Master_Issues tier built. Verify-tag count reduced from 68 to 61, with the remaining 61 now comprehensively tracked rather than scattered. Final state: 47/47 tests passing.

PASS 35 — Claude Sonnet (REF_002_Sources.md comprehensive catch-up audit) — July 2026

Per direct instruction to ensure all verified sources and claims from all documents are recorded in the sources and claims files. Found REF_002_Sources.md had fallen substantially behind the claim ledger's continuous updates across many research passes — everything from the public-space costs, daytime-engagement comparative models, workforce analysis, and the recent citation-correction sweep was missing. Also found and fixed a real duplicate ID (two different sources both labeled "E-001"). Added 40 new detailed source profiles across six organized addenda, cross-checked against the claim ledger systematically (not just spot-checked) across both the most recent research (CL-380+) and a mid-project sample (CL-300-379) to confirm the gap wasn't isolated to only the newest work. Confirmed REF_002_Sources.md's own stated purpose — major, reusable, multi-claim sources with full profiles, not a 1:1 mirror of every ledger row — as the correct completeness bar.

40 new REF_002_Sources.md entries (M/N/O/Q/R series), 1 duplicate ID fixed, REF_002_Sources.md grew from 1,057 to 1,211 lines / from roughly 57 to 97 total source entries. No new claims added this pass — this was a documentation-completeness audit, not a research pass. Final state: 47/47 tests passing.

PASS 36 — Claude Sonnet (second verification round — HMK-012, HMK-003, HMK-004 fully closed) — July 2026

Continued TRACK_001_Master_Issues Tier 6's systematic verification sweep. Closed three full documents this round: HMK-012 (4 items, all confirmed with exact primary-source matches including a precise arithmetic confirmation for RHI Toronto's $609M/1,516-unit figure), HMK-003 (5 items, including a genuinely significant finding — Ontario's own regulation confirms homeless OW/ODSP recipients lose their shelter allowance entirely, with The Trillium supplying the exact dollar figures, $390/month OW and $582/month ODSP), and HMK-004 (1 item, confirmed directly via the official Council record). Found a stronger available quote (Rev. Canon Maggie Helwig) than what the library had been using for the encampment "less visible" framing and incorporated it. Flagged the OW/ODSP finding specifically for future propagation into POLICY-001 given its rhetorical strength, rather than adding it opportunistically mid-verification-sweep.

14 new claims (CL-426 through CL-432, plus renumbering), 11 new REF_002_Sources.md entries, 3 documents fully closed in Tier 6 tracking, TRACK_001_Master_Issues Tier 6 updated to reflect current state. Remaining: HMK-001 (44 items, natural next-round target) and HMK-007 (3 items, correctly gated on legal review). Final state: 47/47 tests passing (pending final stats sync).

PASS 37 — Claude Sonnet (propagation, HMK-001 progress, two queued tasks completed) — July 2026

Per direct instruction: propagated the OW/ODSP shelter-allowance finding into POLICY-001 and DERIV-005 as flagged last round. Continued HMK-001's Tier 6 sweep, closing 3 more items (hotel audit exact citation, heat-days figure advanced with real City data). Completed both queued tasks: (1) audited all ~90 REF_002_Sources.md URLs for paywall-prone academic domains, found and fixed the only two real issues with genuine free alternatives (PMC for Slesnick 2008, the publisher's own free version for O'Grady et al. 2013); (2) re-mined the already-cited TSSS 2025 Budget Notes for additional content, surfacing three new facts — most notably a $7M "year four of a ten-year strategy" already harmonizing POS wages, which meaningfully reframes REC M-11 as completing an existing City commitment rather than proposing something new.

8 new claims (CL-433 through CL-437 plus reused numbering), 2 paywalled links replaced with free alternatives, 3 documents substantively updated (POLICY-001 ×2 recs, HMK-035, HMK-001 ×2 items, DERIV-005). HMK-001's Tier 6 count: 44→42→39 across this and the prior session (net progress this pass: 2 closed + 3 new facts integrated elsewhere). Final state: 47/47 tests passing.

PASS 38 — Claude Sonnet (dedicated HMK-001 round — 13 items closed, 42→31) — July 2026

The dedicated round promised last pass. Worked through HMK-001's remaining ⚠️ VERIFY tags in two phases: quick propagation/consistency fixes first (finding several tags were stale duplicates of facts already confirmed elsewhere in the same document — a real, recurring pattern worth naming, since a large, long-lived document accumulates this kind of internal drift even when the underlying facts are solid), then genuine new research on the harder remaining items (RCMP investigation status, Housing First ROI, precise OW rate history, Canadian CLT census data). Found a genuinely excellent primary source for Ontario social assistance history — John Stapleton, a 28-year former provincial policy official who has written extensively and precisely about exactly this history — worth remembering as a go-to source if the remaining OW-rate-table rows are tackled in a future pass.

7 new claims (CL-435 through CL-441 with renumbering), 4 new REF_002_Sources.md entries, 6 documents/sections updated within HMK-001, TRACK_001_Master_Issues Tier 6 updated to reflect the current, more precisely-categorized remaining 31 items. Final state: 47/47 tests passing.

PASS 39 — Claude Sonnet (continued HMK-001 round — 5 more items, 31→28) — July 2026

Continued the dedicated HMK-001 verification round. Closed the 2013 OW rate row and the CAEH System Leader recommendation with direct primary-source confirmation. Found something more significant while checking the CUPE Local 79 wage claim: it was true when the library was written but has since been resolved by the March 2025 collective agreement, which eliminated minimum-wage jobs citywide. Corrected rather than merely verified — the library now states the historical claim accurately while explicitly flagging it as outdated, and preserves the real, current distinction from REC M-11's actual target population (POS/non-profit contracted operators).

3 new claims (CL-442 through CL-444), 3 new REF_002_Sources.md entries, 3 items closed in HMK-001 including one genuine correction (not just a verification). HMK-001: 31→28 remaining. Final state: 47/47 tests passing.

PASS 40 — Claude Sonnet (OW rate table completed, Helsinki closed) — July 2026

Completed the task explicitly assigned: finished the Ontario Works historical rate table. Found ISAC's own year-by-year percentage archive and used it to reconstruct the full trajectory via compounding from confirmed anchor points. Found and fixed a real error in the process — the 2003 figure was likely wrong (should be $520, unchanged from 1995, given the well-documented Harris-Eves rate freeze), not the higher figure previously used. Advanced 2007 and 2017 honestly, as compounding-derived rather than independently point-sourced. Also closed Helsinki's land ownership/mandate figures via four independent confirming sources.

5 new claims (CL-442 through CL-446 with renumbering), 2 new REF_002_Sources.md entries, the OW rate table fully addressed (a real correction plus two honestly-labeled compounding-based advances), Helsinki closed. HMK-001: 24→23 remaining this session (28→23 across the continuation). Final state: 47/47 tests passing.

PASS 41 — Claude Sonnet (occupancy trend, prevention-cost ratio, new healthcare finding) — July 2026

Continued HMK-001's verification sweep. Advanced the shelter occupancy trend with real City data while explicitly distinguishing fixed historical facts from live data points needing periodic re-checking. Advanced the eviction-prevention-cost ratio via a real comparable (BC Rent Bank's 5:1 ROI) rather than forcing a Toronto-specific figure that doesn't exist in a single source. Found and added a genuinely new fact in the process — a 2024 Unity Health Toronto study's 6x healthcare cost multiplier, methodologically stronger than the library's existing utilization-rate figure.

3 new claims (CL-447 through CL-449), 3 new REF_002_Sources.md entries, 3 items closed in HMK-001, 1 new fact surfaced and integrated. HMK-001: 22→20 remaining this session. Final state: 47/47 tests passing.

PASS 42 — Claude Sonnet (major external research integration — Ombudsman refugee shelter finding) — July 2026

Per direct instruction to integrate user-provided Toronto Council research. Cross-checked and independently re-verified each item before integration rather than accepting externally-compiled research at face value. Found and integrated the most significant new finding in this library's recent history: a December 2024 Ombudsman report finding the City's refugee claimant shelter exclusion amounted to systemic racial discrimination, unactioned by Council, now the subject of active litigation. Added REC M-12, updated HMK-032 with careful framing (strengthens rather than undermines its existing argument), added a council resolution clause, and propagated the recommendation count to 46 across the library. Also integrated a currently-active encampment policy formally opposed by the City's own Housing Rights Advisory Committee, real outreach-model outcome data, and reconciled a minor apparent data discrepancy. Caught two real issues along the way: a false-positive recommendation-count trigger (an unrelated "31" from a different report) and a genuine model-name leakage instance ("Perplexity" in a sourcing note) — both fixed before finalizing.

9 new claims (CL-450 through CL-455 plus supporting), 5 new REF_002_Sources.md entries, 1 new recommendation (REC M-12, 46 total), 4 documents substantively updated (POLICY-001, HMK-032, HMK-003, HMK-001, POLICY_002_Council_Resolutions_Draft, HMK-033), recommendation count propagated library-wide, 2 real bugs caught and fixed (false-positive count trigger, model-name leakage). Final state: 46/47 tests passing (pending final stats sync).

PASS 43 — Claude Sonnet (BRIEF-001 counterweight, HSCIS progress figure updated) — July 2026

Continued propagating the Ombudsman refugee-shelter finding, adding it to BRIEF-001's Council Vote Record section as a necessary counterweight to that section's previously one-sided positive framing of Council's record. Continued mining externally-provided Council records for verifiable value and found a real, current progress update — HSCIS funding at $507.6M/13-of-20 sites (2026 budget), superseding an earlier $258.1M/7-of-20 figure — updated across all three locations where the stale figure appeared, not just the first one found.

1 new claim (CL-456), 1 new REF_002_Sources.md entry, 4 documents updated (BRIEF-001, HMK-012, HMK-029, POLICY-001). Final state: 47/47 tests passing.

PASS 44 — Claude Sonnet (six-document external research batch: verified, recorded, tracked) — July 2026

Per direct instruction to ensure all sources from a large external research batch are recorded with working links, and that sourcing "stands up." Applied full verification discipline rather than bulk-accepting the material, given its scale (six documents, 100+ distinct claims/sources). Checked six specific claims directly: found two unreliable (a false "IHAP 2026 = $0" claim contradicted by the primary budget document; a misattributed "2026 Springer" mortality study whose real figure comes from 2011-2015 Dublin data), and confirmed four (LTB backlog/retention data with a genuinely serious new finding about undisclosed post-tabling data alteration in an official government report; the already-integrated HSCIS progress figure; Built for Zero's methodology update). Built a comprehensive REF_002_Sources.md addendum recording every remaining source category, explicitly labeled by verification status rather than presented as uniformly confirmed. Declined to adopt the batch's proposed new recommendations without independent development. Flagged four specific, high-value items for priority future verification.

5 new claims (CL-457 through CL-461), 1 large new REF_002_Sources.md addendum (7 subsections, R-000 through R-007), 1 new TRACK_001_Master_Issues gap (#45), 3 documents substantively updated (POLICY-001's REC S-09, HMK-009's functional-zero definition). Final state: 47/47 tests passing.

PASS 45 — Claude Sonnet (resource registry built, propagation audited, Finland corrected, external review prepped) — July 2026

Comprehensive four-part response: built a new resource ingestion registry (299+ entries, 7 new automated tests) addressing a real structural gap between the claim ledger and REF_002_Sources.md; audited recent propagation and found HMK-009 had the Finland/Helsinki 2024-2025 reversal fully documented while HMK-001 and HMK-011 were still stale — fixed both, and independently confirmed the reversal continued into 2025 (20% increase, accelerating); found and corrected a real error in the Housing First support-ratio figure (1:3 → the real ACT/ICM standards); wrote a comprehensive new Part 7 addendum to the external review prompt covering everything since Part 6, including honest disclosure of which recent externally-provided claims were verified versus still unreviewed. The test suite itself caught two real issues while this work was underway — both fixed before finalizing, a small demonstration of the discipline working as intended.

6 new claims (CL-457 through CL-463 with earlier numbering), 1 new file (MAIN_DATA_003_Resource_Registry.csv) plus its README, 7 new automated tests, 4 documents substantively corrected (HMK-001, HMK-011, REVIEW_001_External_Review_Prompt, PROC_005_Registry_Readme), 1 comprehensive propagation audit. Final state: 54/54 tests passing.

PASS 46 — Claude Sonnet (queue processing — 5 priority items resolved) — July 2026

Continued systematically processing the resource registry's unreviewed-source queue. All five attempted priority items resolved: the HART Hub CBC investigation (confirmed, integrated, with a real copyright-compliance fix caught along the way — an initial quote draft ran to 29 words and was trimmed to comply with this project's own 15-word hard limit); the Ottawa AG audit (confirmed with an honest date correction, and an honest refusal to integrate an unconfirmed political-response narrative); Diana Chan McNally's candidacy (confirmed, and found already correctly handled from an earlier pass); Finland's 2025 reversal (confirmed, accelerating); and the ICES Cost of Inaction study, which turned out to be more valuable than its original framing — independent cross-validation of an existing finding via a completely different methodology, plus a real new Toronto-specific dollar figure.

5 new claims (CL-464 through CL-467, some renumbering), 4 new REF_002_Sources.md entries, 5 registry entries updated from queue to resolved status, 1 real copyright-compliance catch and fix, 2 documents substantively updated (HMK-001, TRACK_001_Master_Issues gap #45). 18 lower-priority queue items remain, tracked and available for future passes. Final state: 54/54 tests passing (pending final stats sync).

PASS 47 — Claude Sonnet (same-day self-correction: false cross-validation claim caught by registry) — July 2026

While integrating the ICES "Cost of Inaction" study, incorrectly presented it as independent cross-validation of an already-integrated Unity Health finding. They are the same paper (Richard et al., BMC Health Services Research 2024), cited via two different secondary summaries without recognizing the overlap. Caught by the resource registry's own duplicate-URL integrity test — the exact kind of error this tooling was built to catch, working as intended within the same session the error was introduced. Corrected in HMK-001, the claim ledger (both entries annotated and cross-referenced rather than deleted, preserving the audit trail), and the registry (duplicate row removed, surviving entry updated).

0 new claims this pass — pure correction work. 3 files fixed (HMK-001, claim ledger x2 entries, resource registry). Final state: 54/54 tests passing.

PASS 48 — Claude Sonnet (HMK-001 continued, new encampment data flagged not rushed) — July 2026

Continued the Tier 6 sweep. The historical "~60 intake systems" figure couldn't be confirmed in its original form, but found real current value instead — Toronto's STARS system is now a single unified intake tool. Found a new, different encampment count series while researching another item entirely (539→196 vs. the established 283→84) and made the disciplined choice not to merge or integrate it — flagged as a new, explicit reconciliation gap instead, since presenting two different encampment counts as if consistent would be exactly the kind of internal contradiction a hostile fact-check finds first.

2 new claims (CL-468, CL-469), 2 new REF_002_Sources.md entries, 1 new TRACK_001_Master_Issues gap (#46) for the flagged reconciliation, HMK-001 items: 19→17. Final state: 54/54 tests passing.

PASS 49 — Claude Sonnet (encampment metric multiplicity resolved, gap #46 closed) — July 2026

Resolved the encampment count reconciliation flagged last round. Found the City genuinely reports at least four different encampment metrics (locations, people-in-parks, parks-affected, tents/structures) across different contexts — none contradicting the others. This is the right kind of resolution: neither forcing the new figures into false consistency with the existing ones, nor discarding real data because it didn't immediately reconcile. Integrated as a named transparency finding in HMK-003. Also attempted the POS operator percentage figure; found useful context but correctly recognized the precise number remains blocked by the same tooling constraint as an already-tracked gap, rather than forcing an answer.

1 new claim (CL-470), 1 new REF_002_Sources.md entry, 1 gap closed (#46), 1 document updated (HMK-003) with an honest new finding about inconsistent City metric reporting. Final state: 54/54 tests passing.

PASS 50 — Claude Sonnet (hard-tail confirmed unresolved, queue item resolved) — July 2026

Third attempt at the TSSS capital spending rate figure via a targeted primary-document approach; still unresolved, honestly reported rather than forced. Shifted to the resource registry queue and found real value: Feed Ontario's 2025 Hunger Report confirms record Ontario food bank use, and Food Banks Canada's companion report found social assistance recipients now spend 66% of disposable income on housing (up from 49% in 2021) — integrated into REC P-01 as current, striking corroborating evidence.

1 new claim (CL-471), 1 new REF_002_Sources.md entry, 1 document updated (POLICY-001 REC P-01), 1 registry queue item resolved (14 remain). Final state: 54/54 tests passing.

PASS 51 — Claude Sonnet (Denver SIB precise correction, 365→285 housed) — July 2026

Continued the registry queue. Found a genuine correction: the library's Denver SIB "365 housed" figure — itself already the product of one prior correction pass — had never actually been checked against Urban Institute's own evaluation page. The real figure: 363 randomized to treatment, 285 (79%) actually housed. "365" appears to have been a transposition of the randomization count. Corrected with full context. Worth noting: a claim surviving one correction doesn't mean it's been verified against a primary source — this one hadn't been, until now.

2 new claims (CL-472, CL-473), 1 new REF_002_Sources.md entry, 1 document corrected (HMK-001), 1 registry queue item resolved (13 remain). Final state: 54/54 tests passing.

PASS 52 — Claude Sonnet (Chief Coroner queue resolved, new mortality/overdose data integrated) — July 2026

Processed the Chief Coroner registry queue item. Confirmed this library's existing death-count and gendered-mortality figures were already correctly sourced to Toronto Public Health — validating they weren't affected by the Dublin-misattribution issue found in an earlier external-research verification pass. Integrated genuinely new data: a sustained 2026 shelter overdose surge (96/month YTD, roughly double 2025's average) continuing rather than reversing an already-documented trend, and new Ontario-wide context from an ICES study finding shelter opioid deaths more than tripled during the pandemic. Caught and properly consolidated a genuine registry duplicate (not a scripting bug this time — two real claims citing the same source page).

2 new claims (CL-474, CL-475), 3 new REF_002_Sources.md entries, 1 document updated (HMK-021, two additions), 1 registry queue item resolved (12 remain), 1 duplicate properly consolidated. Final state: 54/54 tests passing.

PASS 53 — Claude Sonnet (major substance crisis research: fentanyl, alcohol, stimulants, gendered mortality) — July 2026

Per direct instruction, researched current (especially 2026) data on fentanyl, other addictions, and alcoholism. Major expansion of HMK-021 (four new parts) plus a concise BRIEF-001 addition. Along the way, located the real academic paper behind a statistic an earlier pass had flagged as likely misattributed — the paper is genuine (Ali et al., BMC Public Health, April 2026), and the correction is documented as a clean example of appropriately reasonable-but-premature earlier caution, not as if the original claim had been fabricated. Substantially filled two real gaps in this library's prior coverage: alcohol (30-50% AUD prevalence, Toronto's own Managed Alcohol Program origin story) and stimulants (CAMH's 700%+ methamphetamine ED-visit increase). Flagged two real, evidence-based recommendation candidates (MAP expansion, gender-differentiated shelter capacity) without rushing them into POLICY-001 in the same pass.

9 new claims (CL-476 through CL-484), 7 new REF_002_Sources.md entries, 1 registry ID collision caught and fixed, 3 documents substantively updated (HMK-021 major expansion, BRIEF-001, TRACK_001_Master_Issues new gap #47), 1 propagated minor error corrected (78/85 general-population comparator). Final state: 54/54 tests passing (pending stats sync).

PASS 54 — Claude Sonnet (REC M-13 and M-14 drafted: MAP expansion, gender-differentiated shelter capacity) — July 2026

Developed the two recommendation candidates flagged in the prior pass into fully-scoped POLICY-001 additions. REC M-13 (Managed Alcohol Program capacity expansion) and REC M-14 (gender-differentiated shelter capacity/harm reduction access) both follow the library's standard template and both honestly disclose genuine costing gaps rather than inventing figures. Recommendation count now 48, fully propagated across HMK-033, POLICY_002_Council_Resolutions_Draft, and 10 documents citing the total count. Updated HMK-021's own language to point to the now-drafted recommendations rather than describing them as still-hypothetical.

0 new claims (pure recommendation development, no new factual claims) — 2 new recommendations (M-13, M-14), 48 total. 4 documents updated (POLICY-001, HMK-033, POLICY_002_Council_Resolutions_Draft, HMK-021), 10 documents' recommendation-count references propagated, 1 TRACK_001_Master_Issues gap updated to reflect completion. Final state: 54/54 tests passing.

PASS 55 — Claude Sonnet (Social Planning Toronto resolved, election-timed poverty data integrated) — July 2026

Processed the Social Planning Toronto queue item. Found Toronto is confirmed as having the highest child poverty rate among major Canadian cities (25.7%), from a report explicitly timed to the October 2026 municipal election — integrated into REC P-01. Also found and integrated a real counter-statistic for the HART Hub critique: ~22,000 overdose deaths prevented at Ontario consumption sites before closure. Added immigration-status poverty breakdown to HMK-032 with careful scoping to avoid conflating general poverty data with homelessness-specific data.

3 new claims (CL-485 through CL-487), 2 new REF_002_Sources.md entries, 3 documents updated (HMK-021, POLICY-001, HMK-032), 1 registry queue item resolved (11 remain). Final state: 54/54 tests passing.

PASS 56 — Claude Sonnet (ACTO resolved: Bill 60 upgraded to specific provisions) — July 2026

Processed the ACTO queue item. Upgraded the library's existing general Bill 60 characterization to the four specific concrete provisions (N4 notice halved, N12 compensation removed, 50% arrears payment gate on repair defenses, review window halved), while accurately noting the one reversal (evergreen leases preserved). Also added a real, current counter-example — ACTO's February 2026 court win for a domestic violence survivor's tenancy rights — to keep the tenant-rights coverage from reading as uniformly one-directional.

2 new claims (CL-488, CL-489), 1 new REF_002_Sources.md entry, 1 document upgraded (HMK-003), 1 registry queue item resolved (10 remain). Final state: 54/54 tests passing.

PASS 57 — Claude Sonnet (three more queue items: Wellesley, CCLA, CanLII) — July 2026

Continued autonomous queue processing. Wellesley Institute yielded a real coalition ask (40,000 supportive housing units by 2036, 8 organizations). CCLA yielded a significant escalating story — formal Charter opposition to expanded transit enforcement powers, government proceeding anyway — integrated into HMK-003 connecting to existing Safer Municipalities Act and TTC cost material. CanLII yielded a directly relevant legal precedent (Slapsys v. Abrams) for the Master Lease Strategy's open legal-feasibility question, represented honestly as advancing rather than closing that research item.

3 new claims (CL-490 through CL-492), 3 new REF_002_Sources.md entries, 3 documents updated (POLICY-001/HMK-021 context already covered; HMK-003, HMK-001), 3 registry queue items resolved. Final state: 54/54 tests passing.

PASS 58 — Claude Sonnet (major finding: Waterloo ruling resolves HMK-007's longest-standing gap) — July 2026

Found a landmark May 2026 court ruling while processing the OHRC queue item — encampment clearance without adequate alternatives found to violate Charter rights, confirmed via a joint Federal Housing Advocate/Canadian Human Rights Commission statement. This precisely resolves HMK-007's longest-standing flagged citation gap (the Waterloo precedent, previously unconfirmed for months). Handled with appropriate care given the Tier 0 legal-review gate: the finding advances the citation, but the entry explicitly preserves and reinforces the gate rather than treating "citation found" as "cleared for use." HMK-007's internal VERIFY count is now zero.

1 new claim (CL-493), 1 new REF_002_Sources.md entry, 2 documents updated (HMK-007 in two locations, TRACK_001_Master_Issues Tier 0 tracking), 1 registry queue item resolved (6 remain). Final state: 54/54 tests passing.

PASS 59 — Claude Sonnet (Waterloo case: full examination and comprehensive propagation) — July 2026

Per direct instruction, conducted a full examination of the Waterloo ruling before returning to other work, rather than treating the initial single-source finding as complete. Direct retrieval of the primary decision text revealed a much fuller picture: a two-case lineage (2023 Valente vs. 2026 Gibson, previously conflated), a historic first-of-its-kind Charter finding (homelessness as an analogous ground), a real accountability story (Ford publicly attacking the judge), and — critically — confirmation directly from the government's own appeal announcement that this decision is not final law. Propagated this fully-developed, carefully-caveated picture to every directly relevant document (HMK-007, HMK-003, DERIV-007), while deliberately declining to touch HMK-031 given its explicit high-sensitivity status pending Indigenous organizational review.

4 new claims (CL-494 through CL-497), 3 new REF_002_Sources.md entries, 4 documents substantively updated (HMK-007 twice, HMK-003, DERIV-007), 1 TRACK_001_Master_Issues Tier 0 entry fully rewritten with the appeal-status caveat as its central point. Final state: 54/54 tests passing.

PASS 60 — Claude Sonnet (Sunshine List research: shelter operators and TSSS compensation) — July 2026

Per direct instruction, researched Ontario's Sunshine List for shelter operator and TSSS staff over $100K, integrated into HMK-015 (wage discussion) and DERIV-011 (operator profiles), and attempted FT/PT headcount quantification. Obtained a complete Fred Victor breakdown (24 staff, $3.1M) and TSSS's own GM salary. Confirmed Salvation Army's entry is national, strengthening an existing caution. Could not obtain complete current breakdowns for several other operators due to tooling limitations (disclosed honestly). Critically: could not find reliable FT/PT headcount data for any operator despite genuine effort — this confirms rather than resolves the pre-existing HMK-037/DERIV-011 gap, and was documented as such rather than papered over with imprecise estimates.

4 new claims (CL-498 through CL-501), 4 new REF_002_Sources.md entries, 3 documents updated (DERIV-011 substantially, HMK-015, HMK-037 cross-reference). Final state: 54/54 tests passing.

PASS 61 — Claude Sonnet (Sunshine List second pass: confirmed genuine limit, clarified Salvation Army scoping) — July 2026

Per direct instruction to continue and push further, made a sustained second attempt at Homes First's Sunshine List breakdown (~10 total attempts across both passes), confirming a genuine tooling limitation (JS-rendered data, blocked government portals) rather than insufficient effort — reported honestly rather than either abandoned silently or pursued indefinitely. Added real new value: a full City of Toronto executive breakdown via direct page retrieval, and a second Homes First headcount estimate that conflicts with the first — presented as evidence of aggregator unreliability, not as a resolved figure. Directly clarified for the user that the Salvation Army "restriction" was a scope limitation (national vs. Toronto-specific), not an access restriction.

2 new claims (CL-502, CL-503), 2 new REF_002_Sources.md entries, 1 document refined (DERIV-011) with the fuller, more honest picture. Final state: 54/54 tests passing.

PASS 62 — Claude Sonnet (Sunshine List project close-out: JS-rendering mechanism confirmed and documented) — July 2026

User pointed to a specific Dixon Hall URL believed to show visible data. Direct fetch confirmed the precise mechanism behind this entire task's persistent limitation: individual salary tables on ontariosunshinelist.com are populated by client-side JavaScript after page load, which this library's fetch tooling does not execute — real page metadata (67 records, active years) retrieves successfully, but the specific data tables render empty. This closes out the Sunshine List research task with a precisely diagnosed, rather than ambiguous, final state.

1 new claim (CL-504), 1 new REF_002_Sources.md entry, 1 document finalized (DERIV-011) with the precise technical explanation. Final state: 54/54 tests passing. Sunshine List task closed.

PASS 63 — Claude Sonnet (horizontal review suite — audit only, no fixes applied) — July 2026

Per direct instruction, ran a suite of horizontal (cross-library) reviews to generate a task list, explicitly without executing fixes in this pass. Nine review angles run: cross-reference integrity (HMK/DERIV/REC), terminology consistency (TSSS naming), orphaned-claim methodology check, TRACK_001_Master_Issues queue accuracy, HMK-001 VERIFY-count drift check, recommendation costing coverage, document-count accuracy, test-suite coverage gaps for recently-added critical content, and dangling in-prose "future pass" promises. Findings compiled into a new task list (below/in next user-facing response) rather than fixed immediately. Most checks came back clean or were false alarms from crude matching methodology (documented as such rather than reported as real problems). Two genuinely new, real findings: HMK-012's self-contained "REC H-01 through H-07" tier was never reconciled with POLICY-001's actual recommendation system, and neither the Waterloo appeal-status caveat nor the new Sunshine List figures have regression-test protection against silent future drift.

No claims added, no documents edited in this pass — audit only, per instruction.

PASS 64 — Claude Sonnet (vertical review suite, all 72 documents — audit only, no fixes applied) — July 2026

Per direct instruction, ran document-by-document vertical reviews across all 72 FLAGSHIP files, generating a task list without executing fixes. Full findings delivered to user in-chat. Headline finding: HMK-005 (Shelter Operator Profiles) contains FTE/compensation-band data for exactly the operators this session spent three rounds unable to verify via Sunshine List search — and where directly comparable, HMK-005's Fred Victor figures ($160-200k top band, ~5 staff over $120k) directly contradict this session's own freshly-verified Sunshine List data (CEO $206,371, 24 staff over $100k). This is an internal library contradiction requiring urgent reconciliation, not a new research gap. Also found: MAIN_REF_002_Sources.md has grown to 1,569 lines across 25 unindexed "ADDENDUM" sections with no table of contents; HMK-012 carries an orphaned "REC H-01 through H-07" tier never reconciled with POLICY-001; no regression tests protect several critical recently-added facts (Waterloo appeal status, new Sunshine List figures) from future drift; DERIV-012 has 25 warning/TBD markers including one dangling self-cleanup instruction.

No claims added, no documents edited in this pass — audit only, per instruction.

PASS 65 — Claude Sonnet (major resolution: HMK-005 Tier 0 contradiction resolved via second LLM's data) — July 2026

User routed a second LLM's Sunshine List extraction back into the library. Cross-validated against HMK-005's pre-existing claims: Dixon Hall and YWCA's existing figures confirmed accurate (not fabricated), Fred Victor corrected with an honest explanation (stale data, normal compensation growth — not fabrication), three new operators added (WoodGreen, Neighbourhood Group, Sistering) with complete breakdowns. Definitively resolved the Salvation Army/TSSS department-filtering question (structurally impossible — no such field exists in the disclosure). One surprising finding (Homes First/Covenant House absence from the list) explicitly flagged as unconfirmed rather than accepted, with a specific next verification step named.

8 new claims (CL-505 through CL-512), 2 new REF_002_Sources.md entries, 3 documents substantively updated (HMK-005 to v2.0, DERIV-011, TRACK_001_Master_Issues with 2 new gap entries). Tier 0 item from the prior vertical review closed. Final state: 54/54 tests passing (pending stats sync).

PASS 66 — Claude Sonnet (operator profile expansion: 6 new operators, honest partial coverage) — July 2026

Per direct instruction to expand HMK-005 to the next ten largest operators. Delivered real, verified profiles for six: LOFT Community Services (full financial + multi-year Sunshine List trajectory), Good Shepherd Ministries Toronto and Eva's Initiatives (full financial profiles), plus Na-Me-Res, Sojourn House, and Houselink & Mainstay (confirmed and scale-documented, but no financial profile located — reported as a real gap, not obscured). Caught and prevented a real entity-confusion risk (Good Shepherd Ministries Toronto vs. the much larger, separate Good Shepherd Centre Hamilton). Four target operators not reached in this pass, named explicitly for follow-up rather than silently dropped.

4 new claims (CL-513 through CL-516), 5 new REF_002_Sources.md entries, 1 document substantially expanded (HMK-005 to v2.1). Final state: 54/54 tests passing.

PASS 67 — Claude Sonnet (10-year trend data integrated with discrepancies found and disclosed) — July 2026

User provided historical (2016/2021) Sunshine List data plus a script to pull the raw government CSV directly. Executed the script in-sandbox; confirmed a real network allowlist restriction (proxy-level error, not a data failure) — documented precisely. Cross-checked the new historical data against this library's own prior verification before integrating: found and flagged a ~$56,000 unexplained discrepancy in Dixon Hall's 2025 total between two deliveries from the same tool, and confirmed Neil Hetherington's Dixon Hall CEO tenure was genuine (resolving an entity-confusion concern) while finding his specific cited 2016 salary unconfirmed. Integrated the genuinely valuable 10-year trend pattern (management-tier headcount growth, most clearly at Fred Victor and WoodGreen) with all caveats intact.

4 new claims (CL-517 through CL-520), 3 new REF_002_Sources.md entries, 1 document substantially expanded (HMK-005 to v2.2), 2 TRACK_001_Master_Issues gaps updated/added (#49 strengthened, #50 new). Final state: 54/54 tests passing.

PASS 68 — Claude Sonnet (full integration pass: Sunshine List/operator profiles work verified end-to-end) — July 2026

Per direct instruction, ran a full integration check across every document touched by the Sunshine List/operator profiles work. Found and fixed two genuine staleness gaps: DERIV-011's Sunshine List section only reflected the first expansion round (missing LOFT/Good Shepherd/Eva's, the 10-year trend layer, and the strengthened Homes First/Covenant House finding); HMK-015's wage illustration still described Fred Victor as "the one operator with unusually complete disclosure" when eight operators now have confirmed data. Both corrected to reflect current state. Verified HMK-005's headline funding total ($226M+) against underlying figures — confirmed mathematically accurate. No new claims added — this was a consistency/propagation pass, not new research.

PASS 69 — Claude Sonnet (priorities artifact executed: tests written, 2 more operators, staleness fixed) — July 2026

Worked through the priorities artifact at discretion: wrote both flagged regression tests (Waterloo caveat, Fred Victor stability — suite now 56 tests), added Scott Mission (a genuinely distinct financial pattern worth flagging on its own) and confirmed Margaret's Housing real, found a real partial lead on the Homes First entity-structure question, and updated the priorities artifact itself to reflect current state.

2 new claims (CL-522, CL-523), 3 new REF_002_Sources.md entries, 2 new regression tests, 1 document expanded (HMK-005 to v2.3), priorities artifact updated. Final state: 56/56 tests passing.

PASS 70 — Claude Sonnet (REF_002_Sources.md restructured: comprehensive TOC added) — July 2026

Addressed priorities artifact item #4 (REF_002_Sources.md restructuring). Considered and deliberately rejected a full physical reorganization (moving 1,600+ lines of addendum content into matching A-K categories) as higher-risk than valuable given extensive existing cross-references into specific content throughout the library. Instead added a comprehensive, categorized table of contents grouping all 11 core sections and 25+ addenda by theme (financial/compensation, corrections, legal/rights, population-specific, batch-intake), with explicit line references. Caught my own line-number drift mid-task (edits shift line numbers) and did a verified final correction pass rather than leaving approximate numbers uncorrected — spot-checked two entries against actual file content to confirm exact accuracy.

PASS 71 — Claude Sonnet (Homes First/Dixon Hall comprehensive case studies + verification queue audit, 2 errors caught) — July 2026

Confirmed and acted on a specific user lead: Homes First's audited financial statements disaggregate by individual shelter/property. Direct retrieval of 2017-2020 statements built a genuinely comprehensive case study — per-shelter breakdowns, individual property addresses with mortgage terms, a major growth-trajectory finding (revenue tripled 2017-2020, COVID-hotel-shelter-driven), and a new charitable registration number lead for the standing Sunshine List question. Gave Dixon Hall equivalent depth, with an honest finding that its disclosure structure (program-category, not per-shelter) genuinely differs from Homes First's — stated plainly rather than obscured. Separately audited the library for untracked unverified material, found three gaps (HMK-011, HMK-007/008, HMK-029), and directly executed verification on two HMK-011 claims — both were found to be real errors and corrected (a Houston statistic misread, an uncited Calgary figure).

7 new claims (CL-524 through CL-530), 8 new REF_002_Sources.md entries, 2 documents substantially expanded (HMK-005's Homes First and Dixon Hall sections), 1 document corrected (HMK-011, two errors fixed), TRACK_001_Master_Issues gap #51 added tracking the verification-queue audit. Final state: 56/56 tests passing (pending stats sync).

PASS 72 — Claude Sonnet (Homes First/Dixon Hall accountability research + continued verification queue) — July 2026

Per direct instruction, researched Toronto council records, complaints, wrongdoing, sole-source contracts, and subcontractor information for both flagship case-study operators. Found and carefully reported a real, sensitive finding — a documented 2021 death at a Homes First shelter, handled with full precision per this library's mortality-data standards. Dixon Hall's search came back clean but surfaced a genuine comparative rating finding and the first confirmed subcontractor relationship for either operator (Reconnect Health Services). Sole-source contract data could not be located for either — flagged honestly as a tooling limitation needing MFIPPA/TMMIS access. Continued the verification queue: checked and precisely sourced HMK-007's UN Special Rapporteur citation; checked HMK-007's Kidd 2022 heat mortality figure and could not confirm it — reported honestly as unconfirmed rather than forced to a resolution.

5 new claims (CL-531 through CL-535), 5 new REF_002_Sources.md entries, HMK-005 substantially expanded to v2.4, HMK-007 corrected in 2 places, TRACK_001_Master_Issues gaps #51 updated and #52 added. Final state: 56/56 tests passing (pending stats sync).

PASS 73 — Claude Sonnet (archive system established, Dixon Hall 2024 full precision, honest coverage mapping) — July 2026

Established formal version-archiving practice per direct instruction, following and documenting the project's pre-existing DO_NOT_UPLOAD_SUPERSEDED_* convention. Retrieved Dixon Hall's 2024 statements in full, producing a significant reserves correction ($10.3M Ci figure ≠ $3.6M actual net assets — a real methodological distinction now flagged explicitly), a newly-identified $14.5M unnamed construction project, a disclosed-and-unqualified restatement, and a precise government funding breakdown. Delivered the requested comparative reserves/disclosure-practices analysis between the two flagship operators. Answered "are financials fully analyzed" honestly: no — specific year gaps mapped precisely for both organizations rather than left vague. Advanced (did not close) the sole-source contract question by identifying the correct City document type with a real historical example.

6 new claims (CL-536 through CL-541), 3 new REF_002_Sources.md entries, HMK-005 substantially expanded to v2.5, TRACK_001_Master_Issues gap #52 updated and #53 added, archive system established with README. Final state: 56/56 tests passing (pending stats sync).

PASS 74 — Claude Sonnet (HMK-038 created: historical cost escalation analysis) — July 2026

User asked for a full historical cost-escalation analysis around a striking 2010-vs-current per diem comparison, explicitly framed as a potential "key fundamental part of the story." Researched with full rigor: found the real 2010 system-wide comparator ($52.13, not the single-shelter $56.35 rate originally noticed), computed real inflation-adjusted growth (~96%), and — critically — found and incorporated an honest complication (confirmed 114-125% capacity growth over the same period) that meaningfully revises the user's original "no service growth" framing into a more precise and more defensible claim. Identified and explicitly avoided a real methodological trap (a misleading total-budget comparison that would conflate mandate expansion with cost inflation). Cross-referenced independent corroborating evidence from this library's own Sunshine List research. Fixed an internal HMK-001 inconsistency discovered along the way.

New document created: HMK-038 (v1.0). 4 new claims (CL-542 through CL-545), 3 new REF_002_Sources.md entries, HMK-001 corrected and cross-referenced, TRACK_001_Master_Issues gap #54 added with 5 explicit follow-up items. Final state: 56/56 tests passing.

PASS 75 — Claude Sonnet (three external research sources reviewed, library error caught, HMK-038 expanded to v2.0) — July 2026

Reviewed three independently-provided LLM research outputs on Toronto shelter costs in full before acting, per instruction. Cross-checked the most consequential claims against this library's own existing figures rather than integrating uncritically. Caught and corrected a real, previously undetected error (HMK-026's unconfirmed "$352 hotel shelters" figure, corrected to the AG-confirmed $359 for Warming Centres, with a genuinely separate $253 figure now properly distinguished rather than conflated). Closed a previously-flagged gap (Homes First's 2010 per diem rate, now found at four sites). Substantially expanded HMK-038 with a full 2010 sector-by-sector schedule and an extended 1960-2018 timeline, all clearly labeled by verification status. Flagged one discrepancy (2025 TSSS budget figures) as unresolved rather than picking a source arbitrarily. Identified a convergent, incident-linked data point (COVID hotel rate tied to the already-documented Homes First Novotel death) as meaningfully stronger than an isolated unverified claim.

11 new claims (CL-542 through CL-551, plus CL-546 correction), 2 new REF_002_Sources.md entries, HMK-038 expanded to v2.0 (archived v1.0 first per established practice), HMK-026/HMK-029 corrected, TRACK_001_Master_Issues gaps #54 updated, #55/#56/#57 added. Final state: 56/56 tests passing (pending stats sync).

PASS 76 — Claude Sonnet (comprehensive re-review of all submitted research, HMK-038 rebuilt to v3.0) — July 2026

Per explicit user instruction to take full time and thoroughly re-review all previously-submitted research rather than accept the faster initial pass as adequate. Conducted a complete, methodical extraction across all sources, surfacing a major finding missed the first time: bed turnover collapsed from ~5x/year (2011) to ~2x/year (2024), direct quantified evidence of doubled average length of stay, directly relevant to the document's central open question about service-intensity change. Added the 2009 pre-baseline, the intended-vs-actual provincial funding split, two direct primary-report quotes, four additional historical anchors, and a declining-transparency finding independently confirmed across multiple sources. Explicitly checked and declined to integrate two claims: a "178.3% TSSS increase" figure (resolved via this library's own pre-existing HMK-034 analysis) and a "$75,000/year" figure (traced to a Change.org petition, flagged as weaker sourcing). Rebuilt HMK-038 cleanly to v3.0 given structural drift from incremental edits, with sequential Part 1-9 numbering.

7 new claims (CL-552 through CL-558), 3 new REF_002_Sources.md entries, HMK-038 completely rebuilt to v3.0 (archived v2.0 first), TRACK_001_Master_Issues gap #54 updated and #58 added. Final state: 56/56 tests passing (pending stats sync).

PASS 77 — Claude Sonnet (sourcing audit, RAG-readiness assessment, 2 new automated tests, 1 new audit tool) — July 2026

User directly flagged HMK-038's missing source URLs. Confirmed the complaint (only 2 of 8 source lines had real URLs), fixed it, and traced the root cause to a systemic issue: 25 of 544 claims and 22 of 348 registry entries cite bare-domain placeholder URLs rather than specific documents. Built two new permanent automated tests (11.1, 9.8) tracking this going forward. Assessed broader RAG-readiness: found 31 of 37 HMK documents lack Version metadata (a real gap for future retrieval reliability), and 11 registry entries marked "ingested" with no linked claim. Built and validated a new content-linkage audit tool (properly normalized, distinguishing real gaps from formatting false-negatives) — confirmed a genuine ~62% verbatim-value integration rate, a much more honest metric than either the naive 100% (CL-ID matching) or misleading 10% (unnormalized literal matching) previously reported. Ran a cross-reference integrity sweep (clean, after fixing a bug in the check script itself).

4 new TRACK_001_Master_Issues gaps (#59-61 plus #58 updated), 2 new permanent test suite checks, 1 new standalone audit tool (_tools/content_linkage_audit.py), HMK-038 sources section rebuilt with real URLs and explicit gap-flagging. Final state: 58/58 tests passing.

PASS 78 — Claude Sonnet (HMK-038 finished: bed turnover confirmed via TSSS primary source) — July 2026

Closed the highest-priority open item in the library's cost-escalation research. Direct retrieval of TSSS's own 2026 Budget Notes confirms the bed-turnover finding precisely: 2.02 people served per bed in 2024, tracking began in 2011, continuous decline through 2024 with 2025 the first projected improvement in 14 years. This upgrades the finding from externally-sourced/unverified to primary-confirmed, and is now one of the strongest pieces of evidence in HMK-038 rather than its weakest link. Also precisely confirmed the 2024/2026 per-bed-night gross figures via exact arithmetic from confirmed primary totals, and found a new, real forward-looking risk (TSSS drawing down $272.556M of its $288.902M stabilization reserve in 2026 alone, explicitly flagged by TSSS itself as creating 2027+ funding pressure).

2 new claims (CL-559, CL-560), 1 new REF_002_Sources.md entry, HMK-038 finalized to v3.1, TRACK_001_Master_Issues gap #58 closed. Final state: 58/58 tests passing. HMK-038/per-diem work is now considered complete for this phase.

PASS 79 — Claude Sonnet (final per-diem search + external review prep) — July 2026

Completed the explicit per-diem search: no current official rate found, further confirming the existing declining-transparency finding; found and integrated three new data points along the way (refugee-specific $148/day figure, HSCIS per-bed cross-validation, a flagged $675K/bed outlier-site lead). Then prioritized updating MAIN_REVIEW_001_External_Review_Prompt.md, which was significantly stale, given the user's explicit statement that remaining session time would go to orchestrating external review rather than further research. Added a full new Part 8 addendum covering everything since Part 7 — HMK-005's documented death, the reserves/net-assets correction, the new HMK-038 document and its central unquantified residual claim, and explicit priorities for this review round — plus updated the required-attachment list.

3 new claims (CL-561 through CL-563), HMK-038 Part 9 item 4 closed, REVIEW_001_External_Review_Prompt.md substantially updated with Part 8. Final state: 58/58 tests passing.

PASS 80 — Claude Sonnet (comparative review of second research pass, major AG overpayment finding integrated) — July 2026

User pasted an independently-produced research document covering the same cost-escalation territory and asked for direct comparison and analysis. Cross-checking surfaced a major finding this library had completely missed: a June 2022 Auditor General audit finding $13.2M in documented, contract-violating hotel shelter overpayments (vacancy fees on empty rooms, improper "DMF" charges, meal surcharges) — verified against six independent press sources plus the primary audit document. Integrated into HMK-038 with an explicit methodological note distinguishing this (confirmed contract-administration waste) from the document's main cost-escalation argument (system-level pricing/scale), to avoid conflating two different claims. Also refined the $359 warming-centre figure's framing — it's press coverage's citation of the AG's own $336-491 range, not the AG's literal single number — a precision improvement caught by comparing two independent research passes against each other.

3 new claims (CL-565 through CL-567), HMK-038 to v3.2, 1 new REF_002_Sources.md entry. Final state: 58/58 tests passing.

PASS 81 — Claude Sonnet (multi-source review protocol created) — July 2026

User identified a real, recurring pattern: initial reviews of multi-document pastes skim for headline points and miss substantial material, caught only when explicitly asked to slow down. Diagnosed the mechanism honestly (synthesis competing with extraction for attention within a single read-through, rather than a simple carelessness problem) and built a standing process document separating extraction, cross-comparison, verification triage, and synthesis into mandatory sequential steps. Explicitly noted this is a process commitment, not a testable artifact — no automated check can verify it was followed, unlike the library's content-accuracy tests.

1 new process document (FLAGSHIP_PROCESS_001). No claim ledger changes — this is a self-reflective process artifact, not a content claim.

PASS 82 — Claude Sonnet (memory instruction added, fleet tasks fully covered, intake tracker built) — July 2026

Added a persistent memory instruction (general scope, applies across all future conversations): complete an extraction pass on every source before synthesizing, whenever multiple documents/pieces of feedback are pasted at once. Finished file-upload guidance for the remaining fleet tasks (A, B, D, E, F, G) — all 12 tasks now specify exactly which file(s) to attach at 1/5/10+ file tool limits. Built a new intake-tracking document (MAIN-TRACK-004) mapping every outstanding task and review, logging what's recently closed so incoming results don't re-litigate settled items, and flagging the one thing most likely to need the highest scrutiny when results arrive: any claimed clean current per-diem figure, given two prior thorough searches came up empty.

2 new process documents (FLAGSHIP_PROCESS_001 already existed from prior pass; FLAGSHIP_INTAKE_001 new this pass). File-upload guidance completed for the full 12-task fleet list. No claim ledger changes.

PASS 83 — Claude Sonnet (Task M results processed via new review protocol; DERIV-016 created) — July 2026

First real application of the newly-added memory instruction and PROCESS-001 on a substantial multi-source batch: 8 independent research passes on Task M, one via uploaded PDF. Completed full extraction on all 8 before any synthesis. Verified two high-stakes single-source claims independently before use (the 234,000 Ontario homeless estimate/government disavowal; Ontario's housing-starts ranking, reframed more precisely). Identified and excluded two errors (a garbled Vancouver 2022 account; a likely Rob Ford/Doug Ford name-collision artifact). Caught and visibly corrected a mischaracterization within the same pass: the 234,000 figure was initially described as new when it was already in the library from earlier research — fixed explicitly in the document text, not silently. Flagged a genuine cross-task finding (a 16-3 council vote) that surfaced in the wrong task's results for proper routing to Task H.

4 new claims (CL-568 through CL-571), new document DERIV-016 (v1.0), 2 REF_002_Sources.md entries, intake tracker updated to reflect Task M as processed. Final state: 58/58 tests passing (pending stats sync).

PASS 84 — Claude Sonnet (Task M loose ends closed: MM34.4 vote fully confirmed and routed) — July 2026

Chased down the two items flagged as unverified in DERIV-016. The MM34.4 council vote (encampment shelter-offer limit) was confirmed via direct TMMIS retrieval with a complete, individually-named vote record — far exceeding the original single-line "16-3" claim. Routed to its proper home (DERIV-008) rather than left in the public-opinion document. Found and flagged a genuine connective thread: the same site (66 Third St.) already tracked in DERIV-008 is under active litigation that connects to a separately-flagged HMK-038 cost-outlier lead. Caught and fixed a self-inflicted error mid-pass (an edit that accidentally overwrote TRACK_001_Master_Issues gap #58 instead of adding alongside it).

3 new claims (CL-572 through CL-574), DERIV-008 updated with a fully-sourced vote record, DERIV-016 Part 3 upgraded from unverified to confirmed, 2 new REF_002_Sources.md entries, TRACK_001_Master_Issues gap #62 added (gap #58 restored after a self-caught editing error). Final state: 58/58 tests passing (pending stats sync).

PASS 85 — Claude Sonnet (sourcing audit, BRIEF-001 data loss and full recovery, citation infrastructure built) — July 2026

Measured a real, user-identified gap: 198/199 inline citations across the library lacked clickable URLs despite the backend claim ledger having them recorded. Began remediating BRIEF-001 (highest-traffic document) when a scripting bug emptied the file entirely — no pre-edit archive existed, an established practice not followed here. Recovered full content from the session transcript, rebranded it to current campaign naming, and added clickable URLs throughout. The reconstruction reintroduced five previously-corrected errors (caught immediately by the existing automated regression suite, fixed before presentation — a real demonstration of that suite's value). Built a new permanent test (12.1) tracking inline citation URL coverage as an improving metric going forward. Separately caught and fixed a repeat of an earlier editing mistake (accidentally deleting an existing TRACK_001_Master_Issues entry while adding a new one) — named explicitly as a pattern (gap #65) rather than silently re-fixed.

BRIEF-001 fully reconstructed and remediated (v2.0), 1 new permanent test added, TRACK_001_Master_Issues gaps #63/#64/#65 added (gap #62 restored after a repeat editing error). Final state: 59/59 tests passing.

PASS 86 — Claude Sonnet (BRIEF-001 fully remediated: complete sourcing, further error-hunting) — July 2026

Continued the BRIEF-001 remediation beyond the initial regression-test fixes. Found and fixed an inconsistency the regression suite's regex hadn't caught (a second unhedged COHB causal claim, phrased differently from the first). Traced and verified the $563M ROI figure against HMK-012's own methodology, surfacing an important caveat (the real range is $220M-$620M, with an explicit recommendation to lead conservative under skeptical questioning) that wasn't in the original document and is now added. Hunted down real URLs for 7 more name-only citations via the claim ledger, and independently web-verified 2 more (Latimer et al. 2017's CMAJ Open URL, MHCC's Winnipeg final report) — in the process, independently triple-cross-validating the $58,972 Toronto societal-cost correction and the Latimer 2020 DOI fix made in the prior pass. Final state: 20 of 25 citations directly linked; the remaining 5 are legitimate (an intentional back-reference, and 4 explicitly-marked ⚠️ syntheses pointing to internal methodology rather than a single external source).

BRIEF-001 to v2.1. 2 new REF_002_Sources.md entries. TRACK_001_Master_Issues gap #63 updated to reflect full completion. Final state: 59/59 tests passing.

PASS 87 — Claude Sonnet (HMK-001 and HMK-003 fully remediated, library-wide citation coverage 4%→14%) — July 2026

Continued the systemic citation-URL remediation beyond BRIEF-001, per instruction to "continue fixing everywhere." Worked through the two next-priority documents (HMK-001, HMK-003), matching most citations against the existing claim ledger and independently web-verifying the rest. Found several genuine primary sources in the process: ACTO's Henley Crescent v. Browne Divisional Court ruling, CCLA's formal February 2026 submission and subsequent May 2026 press release on TTC special constable powers, the exact CCPA article underlying the MCCSS funding-cut figures already in the library, the SEIU Healthcare press release behind the Scarborough strike vote, the official Legislative Assembly text for Bill 60, and the City's own CUPE Local 79 ratification release. Checked HMK-012 and POLICY-001 and found they use a different, table-based sourcing style not subject to the same gap — avoided wasted effort applying the wrong fix pattern to documents that didn't need it.

Both HMK-001 and HMK-003 now at or near 100% direct citation linkage. 3 new claims (CL-575 through CL-577), 6 new REF_002_Sources.md entries, TRACK_001_Master_Issues gap #63 updated. Library-wide inline citation coverage improved from 4% to 14% in this pass alone. Final state: 59/59 tests passing.

PASS 88 — Claude Sonnet (work plan built, HMK-013/HMK-003 complete, HMK-002 error corrected, a real gap in the model-leakage test found and fixed) — July 2026

Built MAIN-PROC-004 as a tracked work plan for the full remaining citation-URL backlog (189 citations, 28 documents at start), per instruction to work through systematically until done and to mine each source for additional value while doing so. Completed HMK-013 (15/15) and HMK-003 (fully closed, including 2 additional gaps missed in the prior pass). Made substantial progress on DERIV-009 (9/15 closed).

The most significant finding: while sourcing DERIV-009, found a citation literally attributing a claim to "Claude Sonnet Task E" — an AI model cited as if it were an independent source. Checked why the library's own dedicated regression test for this exact failure mode (7.1) hadn't caught it, and found the test's pattern never included Claude's own name. Fixed the pattern and swept the whole library, finding 8 more live instances across 5 additional documents. One of these led to a substantive content correction: an unsourced "$2.229B/$4.8B by 2035" TCHC backlog figure was traced to a likely digit-transposition of a different real figure entirely, and replaced with the actual confirmed figures from TCHC's own budget notes ($1.752B → $2.540B by 2034).

HMK-013 and HMK-003 fully complete. DERIV-009 at 9/15. 4 new claims (CL-578 through CL-581). 5 new REF_002_Sources.md entries. A real, previously-undetected gap in the model-leakage test fixed, with 9 live corrections resulting. Final state: 59/59 tests passing (pending stats sync).

PASS 88 — Claude Sonnet (systematic work plan launched; AI-model-citation vulnerability found and fixed library-wide) — July 2026

Built MAIN-PROC-004, a formal tracking document for the full remaining citation-URL backlog (189 citations, 28 documents), per instruction to work systematically and mine each source for additional value while fixing links. Completed HMK-013 fully; advanced DERIV-009 to 9/15. The most significant finding: multiple documents cited "Claude Sonnet"/"Sonnet" as if an AI model name were an independent source — exactly the failure mode this library's own model-leakage test (7.1) exists to catch, except the test's pattern never included Claude's own name and so missed it. Fixed the pattern and every live instance, either with a real primary source or honest flagging. One traced citation led to catching and correcting a real content error (TCHC backlog figures, likely a digit-transposition of an unrelated metric). Also found a precise historical correction (the "MOH 2000" document's real title and 1998 publication date) and a real scope distinction in Ottawa's HPP funding figures.

7 new claims (CL-575 through CL-579 already logged, plus 2 more this pass), 1 new process document (PROCESS-002), test_suite.py's model-leakage pattern permanently improved to include Claude/Sonnet, 2 documents (HMK-013, and substantial progress on DERIV-009) advanced. Final state: 59/59 tests passing.

PASS 89 — Claude Sonnet (TCHC correction propagation check: found it had spread, found a second error) — July 2026

Per direct instruction, checked whether the TCHC backlog correction from the prior pass had been fully integrated rather than assuming a single fix was sufficient. Found the identical error had propagated into HMK-004 (with a false ✅), corrected it there too. While re-verifying the full block (not just the flagged figure), found a second uncaught error in the same section — a Facility Condition Index figure with no findable source — and corrected it using TCHC's own primary FCI report and budget notes, which also revealed a more accurate and more useful finding: TCHC's real trajectory (13.4%→14.4%) is moving away from its own stated 10%-by-2027 target. Cross-referenced an already-flagged waitlist reconciliation issue (HMK-025) into HMK-004 for consistency. Mined the FCI source fully rather than stopping at the one number needed, capturing additional structural context about TCHC's portfolio age and declining backlog-specific capital allocation.

3 new claims. Fixed a duplicate-CL-ID bug and a stray empty row in the ledger discovered during this pass. HMK-004 now consistent with HMK-002. Final state: 59/59 tests passing.

PASS 90 — Claude Sonnet (DERIV-009 completed, HMK-035 half-completed, continuing the systematic work plan) — July 2026

Continued the citation remediation work plan. DERIV-009 fully completed (6 remaining citations found and fixed: Medicine Hat's functional-zero achievement, the Safer Municipalities Act's precise statute citation, Bill 28's failure with full co-sponsor detail, a Conversation article with a corrected publication date, plus two ledger-matched fixes). Started HMK-035, completing 7 of 14 citations, including a Lorraine Lam quote from CBC that also surfaced valuable additional data (TTC outreach shelter-referral success rates declining sharply from 2023 to 2024) worth flagging for potential future integration.

5 new claims (CL-585 through CL-588 plus one from the DERIV-009 batch). Work plan (PROCESS-002) updated. Final state: 59/59 tests passing. 4 documents now fully complete (BRIEF-001, HMK-001, HMK-003, HMK-013, DERIV-009); HMK-035 at 50%; ~24 documents remain untouched.

PASS 91 — Claude Sonnet (HMK-035 completed — six documents now fully done) — July 2026

Completed HMK-035 (Public Space Homelessness Costs), closing all 14 originally-flagged citations. Found that the "Cooper/Leading Mobility report" citation was actually the same underlying source (The Local) already linked elsewhere in the document, not a separate unlinked report — checking context before searching avoided an unnecessary search. Found and corrected a genuine source misattribution: a direct quote from TPL's Amanda French had been attributed to the Globe and Mail, but was actually from CBC News (September 16, 2025) — confirmed via the original article and cross-checked against its wide syndication. Flagged one remaining ambiguity honestly rather than resolving it prematurely: the 2026 TPL Program Summary's own snippet shows different cumulative numbers than this document's 2025-specific figures, plausibly reconcilable as different time windows but not confirmed this pass.

2 new claims (CL-589, CL-590). HMK-035 fully complete. Six documents now fully done: BRIEF-001, HMK-001, HMK-003, HMK-013, DERIV-009, HMK-035. Final state: 59/59 tests passing.

PASS 92 — Claude Sonnet (HMK-008 completed — seven documents now fully done) — July 2026

Completed HMK-008 (Federal Funding), closing all 13 originally-flagged citations. Most matched directly against the existing claim ledger. The remaining four (Bill C-12's Royal Assent, the ~30,000-claimant PRRA rollout, and the same-day work-permit mitigation policy) were freshly verified against official Canada.ca and CBC sources, confirming this library's existing figures precisely (the June 3 2025 / June 24 2020 retroactivity dates, the 2-3 day enforcement turnaround, the Minister's same-day signature on the mitigating work-permit policy). One citation (a specific Globe and Mail URL) could not be independently pinned down this pass and was honestly flagged as such rather than guessed.

1 new claim (CL-591). HMK-008 fully complete. Seven documents now fully done: BRIEF-001, HMK-001, HMK-003, HMK-013, DERIV-009, HMK-035, HMK-008. Final state: 59/59 tests passing.

PASS 93 — Claude Sonnet (HMK-021 and HMK-029 completed — nine documents now fully done) — July 2026

Completed both HMK-021 (Mortality Overdose Data, 11/11) and HMK-029 (Full Cost Accounting, 11/11) in a single extended pass. Nearly all citations resolved with real primary sources: ICES/ODPRN's Ontario shelter opioid death study (with named lead author), the April 2026 medetomidine surge in Toronto's drug supply, North Bay Parry Sound's record overdose month (with the previously-uncaptured fatality count), NEJM Catalyst's alcohol use disorder prevalence figure, CATIE's documentation of functional methamphetamine use among unhoused people, Treasury Board's official Value of a Statistical Life figure, AMO's municipal insurance liability data, and the City of Toronto's IPAC-retention commitment during 2023 shelter capacity changes. Every fresh search matched or exceeded the precision already in the library — no errors found in either document this pass, a useful confirmation that not every remaining document holds a hidden mistake.

4 new claims (CL-591 through CL-594 — note CL-591 was HMK-008, included in prior pass log). Nine documents now fully done: BRIEF-001, HMK-001, HMK-003, HMK-013, DERIV-009, HMK-035, HMK-008, HMK-021, HMK-029. Final state: 59/59 tests passing.

PASS 94 — Claude Sonnet (HMK-011 completed — ten documents now fully done) — July 2026

Completed HMK-011 (Political Economy), closing all 10 originally-flagged citations. Confirmed the Calgary Homelessness Foundation's institutional-vs-supportive-housing cost figures exactly via the original literature review. Found real added nuance while verifying London Ontario's veteran functional-zero claim: the achievement required ongoing maintenance and was not perfectly stable, mirroring the Medicine Hat pattern documented elsewhere in this library — added rather than presenting either as a permanent, one-time fix. Confirmed Ontario's internal ministry notes admitting the 1.5-million-homes target was unreachable, with the exact quoted language, replacing a truncated URL with the full correct one.

3 new claims (CL-595 through CL-597). Ten documents now fully done: BRIEF-001, HMK-001, HMK-003, HMK-013, DERIV-009, HMK-035, HMK-008, HMK-021, HMK-029, HMK-011. Final state: 59/59 tests passing.

PASS 95 — Claude Sonnet (HMK-002, HMK-004, HMK-031 completed — thirteen documents now fully done; source-mining applied deliberately per direct instruction) — July 2026

Per explicit instruction to mine sources for additional value rather than just confirm citations, this pass produced real, substantive additions beyond linking:

• HMK-002: found the precise category breakdown behind the October 2025 point-in-time count (9,026 shelter / 1,446 outdoor / 1,724 provincial), that Toronto now runs this count annually under a new federal requirement, and — most valuably — the national Everyone Counts 2025 report confirming the same refugee-claim-driven pattern nationwide, which strengthens rather than merely repeats this library's "not a system improvement story" framing.
• HMK-004: confirmed the "What We Heard" report's full formal citation, authorship, and 176-respondent survey scale.
• HMK-031 (Indigenous Self-Determination Overlay — handled with particular care given subject sensitivity): confirmed TASSC's exact figures, then found and added a second, independently-sourced, peer-reviewed study (Our Health Counts Toronto, n=915) with a corroborating-but-distinct 27.3% figure, presented explicitly as two independent estimates rather than averaged or ranked. Separately, replaced a generic "Wikipedia summary" citation — which this document's own text already flagged as inadequate — with the actual named primary Indigenous critic (Russ Diabo) and his full published argument, directly fulfilling this library's own stated principle that Indigenous critique should be attributed to real voices, not folded into secondary summary. Found and added a direct, previously-absent bridge (via BCCLA) connecting the UNDRIP/self-determination legal argument explicitly to urban Indigenous homelessness — this document's actual subject.

5 new claims (CL-598 through CL-600, CL-603 through CL-605 — with a duplicate-ID fix along the way). Thirteen documents now fully done. Final state: 59/59 tests passing.

PASS 96 — Claude Sonnet (HMK-032 completed — fourteen documents now fully done; deep source-mining continued) — July 2026

Completed HMK-032 (Non-Citizen Populations Federal Funding), closing all 9 citations. Continued the deliberate source-mining discipline from the prior pass with strong results:

• Found that the Ombudsman refugee-shelter Council response included two previously-uncaptured escalation details: this was the first time ever a City Manager refused to endorse Ombudsman recommendations, and Mayor Chow privately expressed regret over how Council handled the vote days later.
• Found a timing nuance resolving an apparent tension in the national-vs-Toronto refugee-shelter-share comparison: the national share kept climbing through 2024 while Toronto's later decline was driven by a documented 2025 national asylum-claim drop — the two datasets are consistent, not contradictory, once dates are aligned.
• Found that international student refusal rates climbed sharply and independently of the permit cap itself (Indian applicants specifically facing ~80% rejection by Q2 2025) — a second, distinct tightening mechanism.
• Found a much larger, potentially significant national figure (300,000+ expired temporary residents) while confirming the PGWP citation — explicitly flagged as unconfirmed rather than asserted, modeling the correct standard of care for an unverified number found along the way.
• The most significant find of the pass: replaced a generic "UN expert statement... cited via Business & Human Rights Resource Centre" citation with the actual primary UN document (A/HRC/57/46/Add.1, para. 22), which states the exact employer-housing-loss mechanism almost verbatim — and separately found that the same UN Special Rapporteur explicitly named "persons experiencing homelessness" alongside Indigenous Peoples as a recognized vulnerability category for exploitation in Canada, a genuinely new and citable connection this library did not previously have.

6 new claims (CL-606 through CL-611). Fourteen documents now fully done: BRIEF-001, HMK-001, HMK-003, HMK-013, DERIV-009, HMK-035, HMK-008, HMK-021, HMK-029, HMK-011, HMK-002, HMK-004, HMK-031, HMK-032. Final state: 59/59 tests passing.

PASS 97 — Claude Sonnet (Full review and integration of Stephen Gaetz / COH research) — July 2026

Per direct instruction, conducted a full review of Stephen Gaetz's body of research (Order of Canada; Director, Canadian Observatory on Homelessness/Homeless Hub; Canada's most-cited homelessness researcher). Checked existing coverage first (one fully-verified paper, credential citations, one bare unsourced bullet) before searching further, to avoid duplicating existing work.

Retrieved and fully reviewed three primary sources: The Real Cost of Homelessness (2012, Gaetz's foundational Canadian cost-of-homelessness synthesis — complementary to, not redundant with, this library's existing Latimer 2020 figures), Without a Home: The National Youth Homelessness Survey (2016, n=1,103, the richest single Canadian youth homelessness dataset reviewed in this project), and the three-generation prevention typology (2017/2018/2024). Created a new document, HMK-039, to hold the full synthesis and catalog lower-priority items for future passes rather than leaving them to be rediscovered.

Made two direct integrations into existing documents: closed a pre-existing bare citation in HMK-010 (the "Roadmap for Prevention" bullet, now properly sourced with the fuller five-point typology and the 2024 current version), and added the National Youth Homelessness Survey's Indigenous-specific national findings to HMK-031 as a third independently-sourced data point alongside the TASSC and Our Health Counts figures already there — explicitly presented as convergent, not averaged or ranked.

Archived both edited files before editing, closing out AI-4 from the July 1 process audit for this specific pass (the standing rule that had lapsed for the entire citation-remediation arc).

4 new claims (CL-614 through CL-617). 1 new document (HMK-039, 80th document in the library). 2 existing documents strengthened with real, properly-sourced additions. Final state: 59/59 tests passing.

PASS 98 — Claude Sonnet (Gaetz/COH review continued; new standing rule added) — July 2026

Added a new standing rule per direct instruction: all unfinished work, open steps, and unexplored ideas must be written down in a persistent tracking location before ending any pass — not left only in conversational text. Added to memory (persists across sessions) and applied immediately within this same pass.

Continued the Gaetz/COH review. Confirmed the previously-flagged, unverified "Can I See Your ID?" (2011) companion criminalization study — 240 youth interviews, real police-contact statistics, and a genuinely new policy recommendation (an amnesty program for SSA-related records) not currently anywhere in this library. Reviewed the State of Homelessness in Canada 2016 report and found its central costed federal ask ($43.788B/10 years, framed as "$50/year, less than $1/week"). Deliberately did not insert either new finding directly into POLICY-001 — flagged both as candidates for a future, properly deliberate addition rather than a same-pass insertion into a 48-recommendation master document.

Closed this pass by writing down, inside HMK-039 itself, both the newly-surfaced open items from this pass and the still-open items carried forward from the July 1 process audit (AI-3 registry backfill, AI-5 log sync) and from the citation-remediation work plan (PROCESS-002's remaining documents, untouched this pass by design since this pass was spent entirely on the Gaetz review per instruction).

2 new claims (CL-618, CL-619). HMK-039 updated to v1.1. Final state: 59/59 tests passing.

PASS 99 — Claude Sonnet (Gaetz/COH review continued, 3 more items) — July 2026

Continued the Gaetz review. Reviewed the LGBTQ2S Youth book chapter (Gaetz 2017) — confirmed its real existence, placement, and the book's own stated framing, but was honest that the chapter's specific policy content could not be retrieved despite repeated searches (search kept surfacing the table of contents and neighbouring chapters, not Chapter 9 itself) rather than inferring content from adjacent material. Flagged a genuinely relevant adjacent lead found in the same search: the same book's Chapter 2, on Indigenous LGBTQ2S youth in BC — not Gaetz-authored, noted as outside this task's literal scope but directly matching an already-flagged HMK-031 gap. Reviewed the Foyer model report (Gaetz & Scott, 2012) and, while searching for it, discovered a previously-unidentified report — Coming of Age (2013) — with confirmed national media coverage across five outlets, flagged for dedicated review given its apparent reach.

Corrected two accidental header-deletion errors made during this pass's own edits (Part 7 and Part 8 headers each briefly dropped when using surrounding text as a str_replace anchor) — caught immediately via structural verification (grep for all ## PART headers) before moving on, not left for a future pass to discover.

Closed the pass by updating HMK-039's own "not yet pursued" list to remove now-resolved items and reflect the current genuinely-open set, and updated the Open Items section accordingly — keeping the standing "write down before closing" rule current rather than stale.

2 new claims (CL-620, CL-621). HMK-039 remains at 80th document, updated in place. Final state: 59/59 tests passing.

PASS 100 — Claude Sonnet (Meta-improvements: two new tools built, largest standing gap resolved) — July 2026

Per a request for meta-level process recommendations, gave direct advice on memory, third-party tooling (Jina, with an honest boundary on what I can invoke directly vs. what belongs in the user's own tooling), and context management (the key point: the files, not the conversation thread, are this project's actual persistent memory — a fresh thread can resume correctly from PROCESS-002/TRACK_001_Master_Issues/HMK-039 alone). Added two new general-purpose memory rules: verify completion computationally rather than from memory, and archive before any destructive edit as a standing reflex.

Then built two reusable tools rather than just describing the fixes: _tools/add_claim.py (eliminates the duplicate-CL-ID bug that recurred three times this session by computing the next ID from the actual file instead of manual tracking) and _tools/backfill_registry.py (resolves AI-3, the largest standing gap from the July 1 process audit — cross-references the full claim ledger against the resource registry and syncs them). Archived the registry before running the backfill. Dry-ran the backfill first to confirm scope (88 URLs) before committing. Ran it for real: 57 new registry entries, 31 existing entries updated with previously-missing claim references. All 59 tests, including the registry-specific integrity checks, passed clean afterward.

Found and logged a new gap while resolving the old one: RAG_SOURCE_EXPORT_priority_ranked.csv is derived from the registry and is now itself out of sync (349 rows vs. the registry's 409) — not actioned this pass, written down per the standing rule rather than left implicit.

No new claims this pass (infrastructure work, not research). Two new reusable scripts added to _tools/. AI-3 resolved; AI-4 now enforced going forward via memory rather than convention. Final state: 59/59 tests passing.

PASS 101 — Claude Sonnet (Projects/file-sharing mechanics researched; RAG export sync closed) — July 2026

Researched Claude.ai Projects and Cowork Projects mechanics against official support docs rather than assuming — confirmed that moving a chat into a project does not automatically share files between chats (requires explicit upload to the project knowledge base), and that Cowork's separate Projects system is the more relevant mechanism for this project's actual working pattern but is local to whichever machine runs it, with no cloud sync. Recommended using the user's already-existing GitHub CLI setup as the actual cross-surface source of truth rather than relying on either Projects system's file-sharing model.

Closed the RAG export sync gap flagged at the end of the prior pass: built _tools/regenerate_rag_export.py, which does a full rebuild from the registry (confirmed derivation logic — priority by claim_count descending, ingestion status from in_sources_md — against the existing file before writing the regeneration script, rather than guessing). Archived the pre-regeneration file first. Registry and RAG export are now both exactly 409 rows.

No new claims this pass (infrastructure work). Three reusable scripts now in _tools/ (add_claim.py, backfill_registry.py, regenerate_rag_export.py). Final state: 59/59 tests passing, registry and RAG export fully synced.

PASS 102 — Claude Sonnet (Projects capacity researched; Gaetz's Coming of Age reviewed) — July 2026

Researched Claude Projects' actual capacity limits against official docs (30MB/file, no file-count cap, 200K-token base context before RAG mode, up to 10x expansion under RAG) and computed this library's real current size (81 files, 2.7MB, ~648K tokens) against them — confirming the library already exceeds the direct-read threshold and would run in RAG/retrieval mode if uploaded to a Project, a materially different and lower-fidelity mode than the direct file access this conversation already has.

Continued the Gaetz review: retrieved and fully read Coming of Age (2014), previously flagged only for its media reach. Found the direct conceptual ancestor of the five-point prevention typology already integrated into HMK-010 (here in an earlier three-part form), a national LGBTQ youth homelessness rate (25-40%) added to HMK-031 alongside TASSC's Toronto-specific Indigenous-youth figure, and a sharp per-diem shelter funding critique — flagged as a new candidate for HMK-029/POLICY-001, not added directly.

First real end-to-end use of the full tool chain built last pass: add_claim.py assigned CL-622 and CL-623 correctly, backfill_registry.py caught the one new URL and updated the one existing entry, regenerate_rag_export.py kept the RAG export in sync automatically. No manual ID tracking, no drift.

2 new claims (CL-622, CL-623). Registry and RAG export remain fully synced (409 rows each). Final state: 59/59 tests passing.

PASS 103 — Claude Sonnet (Gaetz three-lens review: most-cited, Toronto-titled, most recent) — July 2026

Reviewed Gaetz's body of work through three specific lenses per direct instruction. Most-cited: confirmed aggregate citation count (7,184+) but honestly could not surface a clean per-paper ranking this pass — flagged rather than guessed at which single paper is "most cited." Toronto-titled: searching for the 2010 "Surviving Crime and Violence" report instead surfaced a different, earlier, richer 2002 report — Street Justice: Homeless Youth and Access to Justice — a 208-youth Toronto survey providing a genuine 20+ year historical baseline (81.6% crime victimization, family dysfunction and child welfare pathways, landlord/employer exploitation, all with real 2001-2002 figures comparable to this library's more recent data). The originally-targeted 2010 report was not directly fetched — noted as still open, lower priority given the overlap just found. Most recent: found Making the Shift, a $46M national program Gaetz co-led that concluded in March 2025 — current, large-scale, with a specific evidence-based intervention (Family and Natural Supports) not yet reflected anywhere in this library — and a second, distinct 2019 National Youth Homelessness Survey (via Bonakdar et al. 2023) not yet separately reviewed.

Caught and fixed two more header-drop errors during this pass's own edits (Part 8/Part 9 both briefly lost their headers when used as str_replace anchors) — same failure pattern as two passes ago, caught the same way, via structural verification before moving on.

Full tool chain used again for the third time this session: add_claim.py, backfill_registry.py, regenerate_rag_export.py — three new claims, three new registry entries, zero manual tracking, zero drift.

3 new claims (CL-624, CL-625, CL-626). Registry and RAG export remain fully synced (412 rows each). Final state: 59/59 tests passing.

PASS 104 — Claude Sonnet (Fable strategy planning; Gaetz 2019 survey resolved; self-caught stats bug) — July 2026

Advised on whether switching this conversation to Fable retains the work done here (yes — files and environment are tied to the conversation, not the model) versus running 3 external Fable reviews (worse, due to per-conversation file-upload limits that would force curating a subset of this 80+ document library). Recommended switching this thread for one focused task at a time with an explicit scoping instruction, then switching back to Sonnet for routine work, rather than leaving the whole thread on Fable's pricing indefinitely.

Resolved the open question from the prior pass on the 2019 second National Youth Homelessness Survey: it independently replicated most of the 2016 survey's key findings on a larger sample (n=1,375 vs. 1,103) — meaning this library's existing 2016-sourced statistics are validated, not stale, a genuinely reassuring finding worth having on record. Also found a new, recent (2025/2026) racial-disparity finding on Black youth and criminal justice involvement from the same dataset, flagged as a candidate for HMK-001/HMK-031.

Caught and fixed a real self-inflicted bug via the "verify computationally" rule: after adding claims via the tool chain, ran the stats-refresh script but had dropped one of the two regex patterns it's supposed to update, leaving a stale "606 claim ledger entries" line that the test suite correctly flagged. Traced it to the exact line rather than guessing, and found a third, older, differently-formatted stale stats line in the same file while looking — fixed both.

2 new claims (CL-627, CL-628). Registry and RAG export remain synced (414 rows each). One self-caught and corrected stats bug. Final state: 59/59 tests passing.

PASS 105 — Claude Fable 5 (FLAGSHIP_AUDIT_001: whole-library contradiction audit) — July 2, 2026

First Fable pass. Whole-library coherence audit per direct instruction: verified REF_001_Index leads, swept every locked Tier-1 figure for wrong variants, ran automated keyword-clustered dollar-figure collision detection, resolved every flagged cluster. Fixed directly (11 instances, 7 files): the REF_001_Index stats cluster (stale doc count/size, 28/28 test count, stale version footer, stale next-step line — root cause identified: three stats blocks in three formats, only two known to the refresh script); the OW percentage coherence cluster (REF_002_Sources.md sentence citing two different minimum wages in one breath; REF_001_Index row; test 2.2, which was enforcing the INVERSE of verified ledger entry CL-613 — the automated quality gate was defending an internal calc against the sourced figure, and had only ever passed because BRIEF-001's phrasing fell outside the regex window); the $359/$336-491 propagation failure (HMK-038's documented precision refinement had never reached HMK-001 ×2, REF_002_Sources ×2, or HMK-009 — all five fixed, locked with new test 2.22); TRACK_001_Master_Issues's own 45-vs-48 recommendation-count self-contradiction. Also self-caught via test 7.1: my own first REF_001_Index edit leaked a model name into a public doc — removed, recorded in the audit doc rather than silently.

Verified clean (recorded to prevent re-chasing): TCHC/hotel-overcharge wrong variants (all correction-context), $13.8M contexts (Calgary ≠ Toronto), $135 vs $136 (reconciled in REF_002_Sources), $1,058 vs $2,559 per-admission (different studies, correctly scoped), VSL/FCI/Latimer/IHAP/homeless-count families. Flagged for editorial review, not resolved: OW canonicalization across five DERIVs (recommend HMK-001's "about a quarter" rule); DERIV-006's $359 lyric; poverty-industry framing (standing, reaffirmed with locations). Deliverable: MAIN_AUDIT_001_Contradiction_Review.md (81st document). Tests: 59/59 at start, 60/60 at close (test 2.2 rewritten, test 2.22 added). No new claims — this pass corrected internal coherence, it did not ingest new external facts.

PASS 106 — Claude Fable 5 (FLAGSHIP_AUDIT_002: legal/defamation stress-test) — July 2, 2026

Surfaced the existing legal review material (REVIEW_001_External_Review_Prompt §7 — one sentence, explicitly high-level; no dedicated legal prompt existed), wrote the improved dedicated version (MAIN_REVIEW_004_Legal_Stress_Test_Prompt.md — grounded in truth/fair comment/responsible communication/anti-SLAPP, with six priority hunting patterns including AUDIT-001's precision-badge pattern, an anonymity-specific risk analysis, and lawyer-ready output requirements), then executed it against the full library.

Headline: the library's legal discipline is genuinely strong. P1 (named individuals × dishonesty imputations): zero instances. P2 (named operators × impropriety framing): zero adjacency. P4 (unhedged causation-of-death): zero. P6 (litigation/appeal framing): exemplary — the Waterloo under-appeal caveat held in bold at every citation site; the class action is "proposed" everywhere. The standing trafficking boundary held, with HMK-027's treatment a model of the discipline. DERIV-006 already contained its own no-naming guardrail for lyrics. HMK-005 already carried a model compensation legal note. Nothing required unilateral fixing.

What the stress-test produced: (1) a legal read added to the pending poverty-industry editorial decision (innuendo-by-juxtaposition argument; specific hardening steps if kept — anchor to the Gaetz per-diem critique, keep system-level; self-resolving if dropped); (2) five counsel notes recorded in AUDIT-002; (3) the structural finding that the DERIV caveat-inheritance gap is the campaign's largest legal risk — not any single claim — now tracked as new Gap #66, flagged as the highest-leverage pre-publication task; (4) the observation that the campaign's own diligence infrastructure (this log, the test suite, the claim ledger) is itself the responsible-communication defence in artifact form, and should be shown to counsel directly.

Deliverables: MAIN_REVIEW_004_Legal_Stress_Test_Prompt.md and MAIN_AUDIT_002_Legal_Stress_Test.md (82nd and 83rd documents). No new claims — triage and analysis, not new external facts. TRACK_001_Master_Issues updated (poverty-industry legal read, new Gap #66).

PASS 107 — Claude Sonnet (Gap #66 closed: DERIV caveat-inheritance propagation) — July 2, 2026

Back on Sonnet. Closed Gap #66, the highest-leverage item from AUDIT-002. Surveyed all DERIVs against the three specific caveats named, rather than assuming the gap was broad: compensation (DERIV-011 was the only affected DERIV — its header legal note predated the Sunshine List section it now needs to cover; fixed, plus a new safety bullet); $359/range (DERIV-006 is the only real instance, correctly left untouched as the already-flagged pending editorial call — two other apparent matches were false positives on an unrelated figure); Waterloo appeal caveat (DERIV-007 already had it; one apparent match elsewhere was a false positive, a different Waterloo entirely).

Three self-inflicted str_replace anchor-consumption bugs this pass, caught and fixed via the standing verify-computationally rule rather than assumed clean: (1) a duplicate test ID (2.22 used twice) from last pass's edit, never caught until this pass's full-suite run; (2) my own new test's entire host function (test_no_bare_domain_urls) deleted when used as an anchor, caught immediately by a NameError crash; (3) my own new test's record() call deleted in the process of fixing (2), caught not by a crash but by the test silently not appearing in output — required directly calling the function and inspecting the results list rather than trusting the pass/fail summary, then a full systematic sweep (every def test_ function checked for a record() call, not just the one that had just broken) before trusting the suite again.

New regression test added: 10.3, locking the compensation-note propagation as a document-level presence check (not proximity — modeled on the existing Waterloo caveat test, which is the right pattern for a caveat that lives once in a header rather than beside every instance).

No new claims (infrastructure work). Test suite renumbered/expanded: 60→61 tests, all passing. Gap #66 closed.

PASS 108 — Claude Sonnet (PROCESS-002: HMK-005 citation remediation complete) — July 2, 2026

Continuing priorities in order — the three editorial decisions from AUDIT-001/002 are explicitly the user's calls, not mine to resolve unilaterally, so moved to the next item actually available to execute: PROCESS-002's citation queue. Picked HMK-005 (5 gaps, tied for most among untouched documents, and the file I'd just touched for Gap #66 — fresh context).

All 5 gaps resolved directly from the existing claim ledger (CL-346, CL-531, plus the Sunshine List URL already used elsewhere in this library) — no new searches needed, consistent with PROCESS-002's own stated method that most gaps are a linking exercise, not new research. Precisely verified the citation count first via a targeted script (citation-shaped text with no URL within 150 characters) rather than a rough grep, which correctly separated 5 real gaps from one false positive (a bare "(July 2026)" date, not a citation).

15th document fully closed under PROCESS-002. 14 documents remain (~55 citations), next by gap count: HMK-026, HMK-027, HMK-036 (5 each — genuine new-search work likely required, unlike this pass).

PASS 109 — Claude Sonnet (PROCESS-002: HMK-026 citation remediation complete) — July 2, 2026

Second citation document this session. Unlike HMK-005, this one needed genuine new searches — 4 real gaps (the AG report source appeared twice, sharing one URL). All 4 resolved: the AG's warming centre audit backgroundfile; the City's Winter Services Plan retrospective; a TorontoToday article; a Grind Magazine article.

Caught a real ambiguity worth investigating rather than assuming benign: an "atrociously inadequate" SHJN quote attributed to Feb 2026 initially surfaced a near-identical quote from a different SHJN spokesperson dated 2023, about a different winter's plan. Rather than attaching a plausible-sounding URL, searched further and confirmed both quotes (the SHJN characterization and Leslie Gash's "will fill almost immediately") are real, accurately dated to Feb 25, 2026, in a genuine Grind Magazine article — the 2023 instance was simply the same advocacy group reusing similar language in an earlier year, not a citation error.

Reading the actual TorontoToday source (per PROCESS-002's standing method) surfaced real additional detail the library's existing paraphrase had flattened: the original allegation named three sites, not two (Elizabeth St., Willowdale, and Dixon Hall), came from a named advocate (Cathy Crowe, a well-known Toronto street nurse), and a second named source (Greg Cook) was quoted separately calling capacity "massively inadequate." Integrated directly, replacing the vaguer "an outreach worker... two specific warming centres" framing.

16th of 19 PROCESS-002 documents closed. 1 new claim (CL-629). 13 documents remain (~40 citations); next by gap count: HMK-027 and HMK-036 (5 each).

PASS 110 — Claude Sonnet (Process/infrastructure gaps prioritized and closed; 8 optimization recommendations enacted) — July 2, 2026

Prioritized PROCESS-002's own process audit first: verified each AI-item's actual current state computationally rather than trusting the tracker (per the standing rule), and found two were already stale-closed (the RAG export sync, resolved days ago by a tool built after the item was written) while two were genuinely still open (AI-5, AI-6). Closed all four: RAG export sync (verified 415=415), AI-5 (resolved by decision — PROCESS-002 keeps only the status table, narrative logging lives solely in this log, ending the two-log duplication), AI-6 (formalized the internal-cross-reference exemption in writing). AI-4 correctly remains partial — an ongoing reflex isn't retroactively closeable.

Then enacted, not just documented, the 8 optimization recommendations. Built two new reusable tools: check_citations.py (a validated citation-gap scanner — tested against a known-clean and known-dirty document before trusting it, and on that first real test immediately caught 2 genuine gaps in HMK-005 that an earlier narrower manual pass had missed, which is the exact failure mode the recommendation was written to prevent) and close_document.py (a single wrapper combining claim addition, registry sync, RAG export regeneration, a fresh citation scan, and the full test suite into one command instead of several separately-remembered steps). Rewrote PROCESS-002's method section to make source-mining and fresh-scan verification structural requirements, not remembered ones. Documented recommendation #4 honestly as behavioral-only, with no artifact — building a fake tool for a genuine practice reminder would have been worse than naming it plainly.

Fixed the two real HMK-005 gaps found: a Sunshine List citation lacking the direct government URL (only an aggregator was named), and a Charity Intelligence citation for Dixon Hall — which itself surfaced a small new discrepancy (two different Ci charity IDs for Dixon Hall in the ledger), flagged rather than silently resolved, consistent with this document's existing practice for the Homes First CRA# mismatch.

No new external claims (infrastructure and process work). 2 new tools in _tools/, both validated before being trusted. All process/infrastructure gaps from the audit closed. All 8 optimization recommendations enacted with an honest accounting of which were genuine tool-building work versus already-solved versus behavioral-only. Final state: 61/61 tests passing.

PASS 111 — Claude Sonnet (Optimizations fully closed out; six-format stats-drift bug eliminated structurally) — July 2, 2026

Closing out the optimization work from last pass rather than declaring it done prematurely. Found the exact recurring bug AUDIT-001 first diagnosed had already recurred: REF_001_Index's "Automated test suite" line had drifted to a stale 59/59 again. Rather than patch the number a third time, built a self-verifying test (3.6) that checks REF_001_Index's stated count against the true total — placed as the literal last test in the suite specifically so it sees the final count including itself, closing this exact drift class permanently rather than leaving it to recur a fourth time.

Then found the same underlying drift had actually metastasized into six independently-formatted locations across REF_001_Index and TRACK_001_Master_Issues, not the two the earlier audit had caught. Built _tools/refresh_stats.py to consolidate every known format into one script instead of continuing to write inline Python each pass — and validated it via dry-run before trusting it, which immediately caught a real drift point (a "614 claim ledger entries" line still reading 613) that a manual pass moments earlier had missed. Also caught and fixed a bug in the tool's own first draft: the test-suite-count portion of one regex was copying the OLD stale number forward instead of computing a real one, silently reintroducing the exact problem the tool was built to solve. Fixed it by eliminating the redundancy rather than adding more syncing machinery — TRACK_001_Master_Issues no longer independently restates the test count at all, pointing to REF_001_Index's self-verifying copy instead. A sixth drift location (REF_001_Index's DATA-001 catalog entry, stale at "397 claims" against the true 614) was found during a final exhaustive sweep and fixed the same way.

Folded refresh_stats.py into close_document.py as step 4 of what's now a 6-step wrapper, specifically because leaving it as a separately-run script would have recreated the exact "remembered step" anti-pattern recommendation #6 was written to eliminate — caught this before shipping it, not after. Built _tools/README.md, the central tools index that didn't exist before this pass — a real documentation gap for anyone (a new session, a different model, the user themselves) picking this project up without the full conversation history.

No new external claims (infrastructure and documentation work). 2 new tools (refresh_stats.py, _tools/README.md), 1 new test (3.6), close_document.py extended to 6 steps. Six real stats-drift locations found and fixed; one now structurally impossible to recur (test-count, via test 3.6), five reduced from "hope someone remembers" to "one script call." Final state: 62/62 tests passing.

PASS 112 — Claude Sonnet (PROCESS-002: HMK-027 citation remediation complete — first real close_document.py use) — July 2, 2026

Third citation document this session, and the first genuine real-world use of the full close_document.py wrapper (previous runs were validation only). All 5 gaps resolved directly from the ledger — two City Council item URLs, and three sources (CBC on temp-agency wage enforcement, TorontoToday on the YWCA/CUPE 2189 wage dispute, Government of Canada Job Bank) all already present from this document's original research pass, just not surfaced inline. check_citations.py confirmed zero gaps before closing; the wrapper ran registry sync, RAG export, stats refresh, a second citation check, and the full suite in one command with no manual steps in between.

17th of 19 PROCESS-002 documents closed. No new claims (pure citation-linking, no new external facts found). 12 documents remain; next by gap count: HMK-036 (5).

PASS 113 — Claude Sonnet (Source recheck: links verified as accurate, not just resolving) — July 2, 2026

Per direct instruction, went back through this session's citation additions and actually reverified sources — not just that URLs resolve, but that they say what's claimed, distinguishing genuinely fresh reads (HMK-026, done live in Pass 109) from ledger-matched insertions that hadn't been re-read this session (HMK-005, HMK-027).

Found a real error, not just an omission: HMK-005's Dixon Hall Charity Intelligence citation used ID 219, which I had flagged as an unresolved discrepancy against the ledger's second entry (ID 114). Direct verification found 114 is the real, live, current page (2-star rating, B- grade, $26.9M spending — all matching this document's existing figures exactly); 219 was simply wrong. Fixed to the verified-correct ID, replacing the flag with an actual resolution.

Found a genuine staleness issue a working link would never surface: HMK-027's YWCA/CUPE Local 2189 wage-dispute citation was completely accurate as written — every figure (71%/46%/10%, sub-$38,000 pay) checked out precisely against the original May 2025 article — but the dispute itself has since been resolved (ratified May 2026, 11% raise over three years). The citation wasn't wrong; the situation it described had moved on. Updated to present the original figures as historical context for what led to the dispute, with the resolution noted, rather than implying an unresolved present-tense conflict.

Directly verified, word-for-word against primary sources: both City Council procurement items (2024.EC9.4's "10 non-competitive vs. 2 competitive" ratio confirmed exactly; 2023.BA45.8's three named security firms confirmed exactly), the CBC Daisy Warriner story (every detail matches, including the Homes First director's exact quote — and confirmed the library's existing choice not to include her mother's separate, more legally complex hospital-restraint death was the right editorial restraint, not an oversight), and the Job Bank wage range ($19.79–$38.00/hour, confirmed current).

No new claims — this was verification work on existing citations, not new research. 2 real issues found and fixed (a wrong Ci ID, one staleness update) across HMK-005 and HMK-027. 8 sources reverified as accurate, not merely functional. Final state: 62/62 tests passing.

PASS 114 — Claude Sonnet (Full-library review for a prioritized next-steps list) — July 2, 2026

Reviewed every tracking mechanism computationally rather than from memory, per direct request: TRACK_001_Master_Issues (all 7 tiers plus pending-decisions and strategic-observations sections read in full), PROCESS-002's live status table, HMK-039's own open-items tail, and a library-wide sweep for any other document carrying its own open-items section (none found beyond those two). Found and fixed one real drift while reviewing: HMK-039's tail section still listed AI-3 and AI-5 as open (both closed in Pass 110) and described HMK-005/026/027 as untouched (all three closed since). Corrected in place rather than left to mislead whichever session reads it next — exactly the kind of cross-document staleness this whole review was meant to surface.

No new claims — a synthesis and one staleness fix, not new research. Full prioritized next-steps list delivered directly to the user rather than only logged here.

PASS 115 — Claude Sonnet (PROCESS-002 CITATION QUEUE FULLY CLOSED — all 19 documents) — July 2, 2026

Finished the entire remaining citation queue in one continuous pass: HMK-036, DERIV-004, HMK-007, HMK-025, HMK-030, DERIV-007, DERIV-011, HMK-017, HMK-024, HMK-019, HMK-038, HMK-015 — 12 documents, closing out from the 7 already done. Every one followed the established method: check_citations.py first, ledger check second, live search only for genuine gaps, close_document.py to finish.

Real findings along the way, not just mechanical linking: DERIV-004 turned out to have zero real gaps — its "citations" were FOI-letter template placeholders ([DATE] fields), a tool-pattern false positive correctly identified rather than force-fitted with fake URLs. HMK-024 similarly resolved to zero real gaps — its one apparent citation was a legitimate internal cross-reference to HMK-001, which already carries the real URL. HMK-007's legal citations required finding primary CanLII/court-decision sources rather than secondary summaries — direct PDFs for both Waterloo cases now cited. HMK-025 caught its own tool: a citation that passed check_citations.py only by accidental proximity to a different citation's URL a few lines above, not because it had one of its own — fixed with a real URL rather than trusted on a technicality. HMK-019 and HMK-038's OACAS and 2009 Staff Report figures were both confirmed exact-match against primary sources before linking.

No new claims — pure citation-linking and verification work across this batch, consistent with most of these documents already containing their substantive facts from earlier research passes. PROCESS-002 marked COMPLETE in its own header. All 19 originally-tracked documents closed; zero "Not started" entries remain. Final state: 62/62 tests passing.

PASS 116 — Claude Sonnet (Tier 3 closed: 3 of 3 actionable items, all already-resolved-but-untracked) — July 2, 2026

Worked through TRACK_001_Master_Issues's Tier 3 as directed. Of five items, #14 is correctly blocked on a real calendar date (Aug 21 nomination close) and #27 was already marked mostly-resolved — leaving #25, #26, #28 as the three actionable items, a clean match for "3 items."

All three turned out to be the same finding: genuinely resolved elsewhere in the library, with only this specific tracking table never updated to say so. Gap #25 was half-true — HMK-001's own text already claimed the reconciliation closed it, but HMK-003 carried no reciprocal cross-reference, so the fix was only one-sided; completed it by adding the matching note to HMK-003 §G. Gap #26 was already fully resolved — REC 15's descriptive text carries a caveat more prominent than the one the gap complained was missing, naming three specific unanswered legal questions directly, with real supporting case-law research added since. Gap #28 was already resolved — Tier 0 item 3 already names FOI-004 individually and says so in its own text.

Rather than fix three lines quietly, named the pattern explicitly in the tracker itself: three-for-three "already done, tracker just not updated" is a signal worth surfacing, not hiding — the actual fix (closing work) and the tracking fix (updating TRACK_001_Master_Issues) are two separate steps, and the second keeps being skipped. Added this as an explicit standing note in Tier 3 rather than letting a future pass rediscover the same pattern a fourth time.

No new claims — synthesis, one substantive addition (HMK-003's cross-reference), and tracker accuracy work. Final state: 62/62 tests passing.

PASS 117 — Claude Sonnet (Returning to Gaetz: 3 integration items closed, plus a real copyright fix found along the way) — July 2, 2026

Both prerequisite tasks (citation queue, Tier 3) were complete, so returned to Gaetz as explicitly instructed rather than opening a new thread. Integrated three of five outstanding candidates from HMK-039's backlog, each into its correctly-matched home document: Making the Shift's "Family and Natural Supports" finding into HMK-010's five-point prevention typology (the current evidence base for the "early intervention" category); the per-diem funding critique into HMK-015 §1.1, framed as independent national academic corroboration of the Toronto-specific AG finding already there; the Black youth CJS involvement finding into HMK-001 §23, explicitly distinguished from — not merged with — the existing Toronto-specific OSSA ticketing finding, since they measure different things at different scales.

A real compliance issue found and fixed while doing this work, not after: the per-diem integration's first draft quoted Gaetz's exact sentence verbatim — well over the 15-word copyright limit. Checking whether this was a new mistake or a repeat found the same 40+-word verbatim quote had existed in HMK-039 since Pass 102, undetected. Both instances paraphrased down to a single short, protected phrase ("rewarded for keeping people homeless"), with a new regression test (2.24) locking the fix so it can't silently reappear. A full-library sweep confirmed no other instances of this specific quote existed.

Two integration candidates remain open (the amnesty-program recommendation, the $43.788B federal-ask framing device), plus the full Part 8B source-review list — written down in HMK-039's own tracker, not left only here.

3 new claims (CL-630 through CL-632, added via the full toolchain). 1 new regression test (2.24, copyright fix). Registry, RAG export, and stats all synced automatically. Final state: 63/63 tests passing.

PASS 118 — Claude Sonnet (Copyright clarification; REC M-15 added — 4th Gaetz backlog item closed) — July 2, 2026

Answered a direct question honestly: the 15-word quotation limit is not Canadian copyright law (which uses a multi-factor fair dealing test with no bright-line word count, per CCH Canadian v. Law Society of Upper Canada) — it's a conservative operating constraint applied regardless of jurisdiction, flagged as narrower than what Canadian fair dealing would likely permit for a research/criticism library, with the actual legal ceiling left to the queued lawyer review rather than self-assessed.

Then closed the fourth of five Gaetz backlog items: added REC M-15 (municipal Safe Streets Act record amnesty) to POLICY-001, matching the exact M-13/M-14 format — evidence paragraph, bolded ask, Evidence/Authority/Timeline/Cost line. Genuinely the lowest-cost recommendation in the full set (administrative only, no capital or ongoing budget), which made its Cost line unusually easy to write honestly compared to M-13/M-14's disclosed gaps.

Propagated the resulting 48→49 recommendation count fully: a new council motion (clause 26) matching the existing format, a new HMK-033 cost-benefit matrix row, and 11 stale "48 recommendations" mentions fixed across 10 files (MAIN_REF_001_Index, POLICY_002_Council_Resolutions_Draft, REVIEW_001_External_Review_Prompt, REF_004_First_100_Days, HMK-010, HMK-012, HMK-019, HMK-033, TASK_001_LLM_Fleet_Tasks, REVIEW_003_Review_Menu). Test 8.1's ground-truth count is computed dynamically from POLICY-001's own REC tags, so it caught every propagation gap without needing a manual update.

1 new claim (CL-633). 1 new POLICY-001 recommendation. Registry, RAG export, and stats synced via the toolchain. 4 of 5 Gaetz integration candidates now closed — only the $43.788B federal-ask framing device remains from the original backlog, plus the full Part 8B source list. Final state: 63/63 tests passing.

PASS 119 — Claude Sonnet (Final Gaetz integration candidate closed — all 5 done) — July 2, 2026

Closed the last of the five original Gaetz integration candidates: the $43.788B/10-year national federal ask with its "$50/Canadian/year, less than $1/week" framing device, added to HMK-008's federal asks section as an explicitly-scoped rhetorical tool for the scale of the itemized City-specific asks, not a figure requiring reconciliation against them.

Caught a real discrepancy before finalizing rather than after: the URL I recalled from memory for this source didn't match the path already verified in the claim ledger (a /sites/default/files/ vs /wp-content/uploads/2023/12/ difference — the same PDF filename, two different URL structures, only one previously confirmed). Checked the ledger before trusting memory, per the standing rule, and used the already-verified path.

All five original Gaetz integration candidates are now closed. The only genuinely open Gaetz thread remaining is Part 8B's source-review list (LGBTQ2S chapter content, the Indigenous LGBTQ2S BC chapter, What Would It Take?, Making the Prevention of Homelessness a Priority, and the 2013/2014 State of Homelessness editions) — new-source reading, not integration of already-found material.

1 new claim (CL-634). Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 120 — Claude Sonnet (Dedicated Gaetz search: Toronto-specific and foundational work) — July 2, 2026

Per direct instruction, ran a fresh search pass specifically targeting Gaetz's most important and Toronto-focused work, rather than continuing from the existing backlog. Found genuinely new material across multiple angles: The Missing Link (2006, with O'Grady) — the primary Gaetz-authored source underlying the "systems prevention" category this library's five-point typology already relies on, never directly cited until now; an independent academic finding (not Gaetz's own) showing OSSA ticket enforcement is heavily concentrated — 6.2% of ticketed individuals received 51.4% of all tickets — a real strengthening addition to the existing HMK-001 criminalization critique; the State of Homelessness in Canada 2014 edition, now properly distinguished from the 2013 edition (previously conflated as one unreviewed item), with real, previously-uncaptured figures ($7B/year cost vs. $119M/year federal response). Also confirmed several additional real Gaetz titles (Safe Streets for Whom? 2004 — likely the intellectual root of the OSSA critique this library already leans on heavily; O'Grady & Gaetz on gender/income generation; Buccieri & Gaetz on H1N1 as a pre-COVID pandemic-preparedness parallel) catalogued for a future pass rather than chased individually given time constraints.

Caught and fixed two more header-drop errors during this pass's own edits — the same recurring failure mode, same fix pattern (verify via grep "^## PART" immediately after editing, restore before moving on). This time the fix required a genuine renumbering (8B/8C) rather than a simple restoration, since the original single insertion attempt left one section's content orphaned with no header at all — caught by checking the actual line content around where a header should have been, not just checking that a header existed somewhere.

3 new claims (CL-635 through CL-637). Three new integration candidates flagged for a future pass, not added directly — consistent with this session's established practice of separating "find and verify" from "integrate," especially for a fresh search pass rather than a planned integration session. Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 121 — Claude Sonnet (Second search pass: pandemic-preparedness find; three prior candidates integrated) — July 2, 2026

Second dedicated search pass found genuinely new material: Gaetz & Buccieri's "The Worst of Times" (2016) — a chapter on pandemic planning and homelessness, written four years before COVID, warning specifically about shelter crowding as infection risk and policing responses to homeless populations under public-health pretext. Full text retrieved and read. Not integrated this pass — this library has no existing pandemic-preparedness section for it to extend, and forcing it into an unrelated structure would be worse than leaving it properly flagged for a dedicated future addition. Also partially resolved the standing mystery of the recurring 2010 Gaetz quote (likely a direct predecessor to The Missing Link, specific title still unconfirmed) and identified a book-chapter version of the OSSA research not yet compared against what's already cited.

Then moved into integration as instructed, closing all three candidates flagged at the end of the prior pass: The Missing Link (2006) added to HMK-013's corrections-discharge section as historical corroboration — showing the John Howard Society documented the identical structural failure nearly twenty years before the current 2023-2025 data, sharpening the accountability argument considerably. The independent ticket-concentration finding (6.2% of ticketed individuals received 51.4% of all tickets) added to HMK-001 §23 as a third dimension of the existing OSSA critique, alongside the Toronto enforcement-volume and national racial-disparity findings already there. State of Homelessness 2014's $7B/year cost vs. $119M/year federal investment ratio added to HMK-012 Part 1, explicitly scoped as non-additive national historical context rather than a Toronto-specific figure.

3 new claims (CL-638 through CL-640). Registry, RAG export, and stats synced via the toolchain. One new high-value finding flagged, honestly, as not yet having a clean home rather than forced into one. Final state: 63/63 tests passing.

PASS 122 — Claude Sonnet (Third Gaetz search: a major cross-government accountability finding; the recurring header bug finally gets a real test) — July 2, 2026

Third dedicated search pass. Found the most significant single finding of this whole Gaetz thread: Gaetz was a formally-named member (confirmed directly from the primary report) of Ontario's 2015 Expert Advisory Panel on Homelessness, whose central recommendation the Wynne government formally accepted — end chronic homelessness within 10 years, a 2025 deadline. Integrated into HMK-003 §E with deliberately careful, explicit framing: not attributed to Ford, who took office three years after the target was set, but presented as an inherited benchmark his government governed against for roughly six of the target's ten years, expiring with conditions documented elsewhere in this library as dramatically worse, not better. Cross-checked the cited Toronto figures (7,300→15,418) directly against DERIV-007 before using them rather than trusting memory. Also confirmed Gaetz's parallel 2017 federal advisory role, sharpening how his other work in this library should be read — not an outside critic, a formally-appointed advisor at both levels of government this campaign holds accountable.

The recurring header-drop bug happened a fourth time during this pass's own edits — same failure mode, same file (HMK-039), caught the same way (manual grep before presenting). Given four occurrences of an identical failure with no automated test ever catching it, built one: test 10.4 locks HMK-039's full Part-header sequence and fails the suite if any are missing, closing the gap between "I always catch this manually" and "the suite catches it whether or not I remember to check."

2 new claims (CL-641, CL-642). One new regression test (10.4) targeting a demonstrated, repeated failure point rather than a hypothetical one. Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 123 — Claude Sonnet (Task A returns processed: a caught contradiction, a major census development) — July 2, 2026

Received two Task A submissions (external LLM research on Toronto-specific hidden-homelessness measurement) — noted honestly that a third "pasted" submission the user referenced did not come through in what was received, rather than proceeding as if three were present. Confirmed the second uploaded PDF was a duplicate of the first (identical MD5 hash), not a distinct third piece.

The two genuine submissions reached the same ultimate conclusion (2021 Census PUMF cannot produce a measured Toronto-specific hidden-homelessness figure; restricted access required) but directly contradicted each other on a load-bearing factual point: whether PUMF geography reaches the census-tract level. Rather than average, pick one, or present both without resolution, checked directly against Statistics Canada's own official product catalogue page — confirmed geography is CMA-level only, resolving the contradiction decisively and identifying a real, meaningful error in one of the two submissions.

While verifying, found something that substantially changes this gap's status: Canada's 2026 Census, conducted in May 2026 — two months before this writing — included dedicated hidden-homelessness questions for the first time in Canadian census history, following years of advocacy and a 2024 test involving 222,000 households. This is expected to produce the first-ever national measured hidden-homelessness estimate, honestly caveated (private dwellings only, no confirmed release date found). Updated TRACK_001_Master_Issues Gap #6 from "blocked, no path forward" to "blocked on a known, dated instrument with an unconfirmed release timeline" and closed out Task A in the fleet-tasks document with an explicit recommendation: don't pursue restricted 2021 data access, monitor for the 2026 Census release instead, since it will be a structurally better answer than anything achievable with the outgoing data.

2 new claims (CL-645, CL-646). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 124 — Claude Sonnet (NEW DOCUMENT: HMK-040, comprehensive legal/international standards review) — July 2, 2026

Built a new, comprehensive document per direct instruction to review Canadian legislation, international standards, UN research, and SPHERE standards "exhaustively and deep." Checked HMK-007's existing coverage first to avoid duplication (Charter, OHRC, NHSA, Special Rapporteur general content, case law all already there) before researching genuinely new ground.

Major new findings, each independently verified against primary sources: Canada has been formally criticized by UN treaty bodies across at least four separate reviews spanning 1993-2016 (CESCR three times, the Human Rights Committee once in 1999 — the latter explicitly linking homelessness to death, a quarter-century before this library's own mortality data does the same for present-day Toronto). A UN Special Rapporteur physically visited Toronto as part of an official 2007 fact-finding mission — not a report written from Geneva, an in-person visit with recommendations matching this campaign's asks almost word for word, delivered twelve years before Canada acted on them via the NHSA. The Federal Housing Advocate — a real Canadian statutory oversight body, not an international one — produced a 2024 encampments report and a very current 2026 follow-up with devastating national figures (107% rise in unsheltered homelessness, a named and dated Indigenous death in a tent fire, a Toronto mortality figure flagged as unverified rather than trusted secondhand). Ontario's Housing Services Act, 2011 was traced directly to confirm HousingTO isn't merely a policy choice — s.6 makes a housing and homelessness plan a legal requirement for Toronto specifically.

On SPHERE specifically: verified the actual standard (3.5m² warm climate / 4.5-5.5m² cold climate per person) against Toronto's own shelter standards (3.5-3.75m²) independently, rather than asserting relevance abstractly. Found Toronto's shelter space matches SPHERE's warm-climate humanitarian minimum despite Canada's climate, and — a stronger, entirely self-referential finding requiring no international-law appeal at all — is below what Toronto's own Property Standards Bylaw requires private landlords to provide ordinary tenants (4m²/person). Explicitly framed SPHERE's actual scope honestly (humanitarian/disaster response, not binding on domestic social services) rather than overclaiming legal force.

5 new claims (CL-647 through CL-651). One new document (HMK-040, the 84th in the library), cross-referenced from HMK-007. New Gap #67 tracks the real open threads (an unverified mortality figure, an unverified Ontario-wide figure, an unverified prison-standard comparison, an unintegrated $250M federal program). Final state: 64/64 tests passing.

PASS 125 — Claude Sonnet (Pulling threads: mortality figure traced and corrected, UHEI integrated) — July 2, 2026

Worked through the open items flagged at the end of HMK-040's creation, per direct instruction to continue pulling threads rather than leave them listed and untouched.

The mortality figure: traced the Federal Housing Advocate's "59 deaths, Toronto 2025" citation to its actual primary source and found a real error, not a confirmation — the number is genuine but belongs to 2024, not 2025, and to TSSS's narrower "shelter residents" category (the same scope as the library's already-verified 132-in-2021 figure), not the comprehensive "all people experiencing homelessness" total the library otherwise leads with (331 for 2022, 300 for 2023). Corrected in HMK-021 with the distinction stated precisely, and in HMK-040. The genuine comprehensive 2024/2025 total remains unfound and is now the real open item, correctly distinguished from the now-resolved shelter-resident figure.

The UHEI program: found Toronto's exact bilateral federal allocation ($25,799,702 over 2024-25 and 2025-26, cross-verified across three independent sources) and expanded what had been a thin, buried line item in HMK-008 into a full standalone entry — with the real finding being the funding ratio itself: Toronto's own $400M contribution against $25.8M federal, about 15.5 times the federal amount, alongside the Federal Housing Advocate's own December 2025 assessment that the program's short duration undercuts its impact. While researching UHEI's Toronto-specific documentation, found a genuine bonus: a more current Central Intake figure (216 unmatched callers/day, Nov 2024) extending the library's existing 174→202 trend rather than contradicting it — added to HMK-024.

Two of four originally-flagged threads closed; three lower-priority ones from HMK-040's own list (federal prison standard, Protecting Tenants Act 2020, primary-form 1993/1998 CESCR documents) and one newly-flagged one (Ontario's 85,000 figure needing a consistency check) remain genuinely open, written down rather than left implicit.

3 new claims (CL-652 through CL-654). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 126 — Claude Sonnet (DeepSeek report reviewed: 2 significant errors caught, 4 real additions integrated) — July 2, 2026

Received a substantial external research report and checked its highest-stakes claims against primary sources individually before integrating anything, rather than trusting its polished presentation. Found two significant errors and confirmed several genuine additions.

The major catch: the report presented Ontario's "Homelessness Ends with Housing Act, 2025 (Bill 28)" in present-tense authoritative language, as if it were enacted law. Direct verification against the Legislative Assembly's own bill tracker and Hansard found it is a private member's bill from opposition MPPs that was defeated at second reading on October 22, 2025 — the Ford government's own side voted it down. Corrected in HMK-040 with the accurate story, which is arguably more useful to the campaign than the false one: a broadly advocate-endorsed 10-year homelessness-elimination framework, killed by the government this campaign is holding accountable, is a stronger and entirely real finding. Flagged as a strong integration candidate for HMK-003 §E.

The second catch: a specific OECD cost-benefit figure ("CAD 2.20 saved per dollar") did not match the OECD's own primary text, which states USD 1.44, specific to the US. Corrected in HMK-040 §6.6.

Confirmed accurate and integrated: Bill C-205 (federal, real, but independently found to be stalled — a detail the source report omitted); Bill 6/Safer Municipalities Act (largely accurate, with additional real detail found during verification — 703 encampments, the 74-39 vote, and a striking Toronto-specific Indigenous statistic from Yellowhead Institute flagged for HMK-031); the Toronto Housing Charter (genuinely new, verified, a real municipal rights document not previously in this library).

Four claims left entirely unverified and not integrated anywhere (Build Canada Homes Act, Homelessness Task Force Act/Bill 204, the Canadian Centre for Housing Rights' 2024 shelter standards, two named Toronto documents) given the report's demonstrated error rate on its most load-bearing claims — flagged as leads requiring independent verification, not treated as fact.

A genuine self-inflicted issue caught before it reached the user: while writing up this review, briefly named the source model ("DeepSeek") in a new section header — exactly the failure mode test 7.1 exists to catch. Caught it via a full test run before presenting, scrubbed it, and — while checking for other instances — found this exposed a real, separate, pre-existing gap: several older model-name leaks in HMK-001 and POLICY-001 (Gemini Free Thinking v2, Qwen37pT, Qwen37plus2) that survived a prior session's ~130-instance cleanup and that test 7.1's current pattern list doesn't catch. Logged as new Gap #68 rather than silently expanded into today's scope — a real fix, not yet done.

Also caught and fixed the recurring section-header-drop bug twice more during this pass's own edits (sixth and seventh occurrences this session), both caught immediately via the now-standard practice of grepping headers right after each edit near a boundary.

4 new claims (CL-655 through CL-658). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 127 — Claude Sonnet (Qwen report reviewed: a more serious error found; to-do audit; two new tracked projects) — July 2, 2026

Applied the same verification discipline established last pass to a second external report on the same subject. Found a more consequential error than the first one: the report described a "Homelessness Ending Plan" as official City of Toronto policy, with a specific quantified claim (75% of shelters, 90 of 120, converted to permanent housing) and a cited toronto.ca URL. Direct retrieval of that exact URL found it is not a City document at all — it's a private citizen's unsolicited personal budget proposal ("Toronto HomeFirst," submitted by James D. Golding, a homelessness survivor advocating for his own plan), filed as one of many public submissions during 2026 budget consultations. This is a more serious failure mode than the prior pass's Bill 28 error, since Bill 28 was at least real legislation; this was mistaking one citizen's advocacy document for institutional policy. Also checked a UN "Domicide" report the source cited — confirmed real (A/HRC/61/43/Add.3, Feb 2026) but concerning active armed conflict (Gaza, Myanmar, Sudan, Ukraine), a category mismatch rather than a fabrication, not integrated on that basis. Both findings written up in HMK-040 Part 8.2, alongside the genuinely verified new material (matches Part 8.1's already-confirmed items).

Conducted a full to-do audit across HMK-039, HMK-040, and TRACK_001_Master_Issues, finding several items from the Qwen review that hadn't yet been added to any tracked list (Bill 60, the "four populations by 2025" claim, the Protect Ontario by Building Faster Act, a SPHERE urban-settings companion document, and a document-dating discrepancy) — added all of them explicitly rather than letting them exist only in this response's own text.

Set up the two new requested projects. The systemic readability/flow review was logged in TRACK_001_Master_Issues's Strategic Observations as explicitly deferred, per direct instruction, with a note on why (several documents are still being actively extended and a restructuring pass now would likely need redoing). The source-discovery prompt was designed as a new Task N in TASK_001_LLM_Fleet_Tasks.md, following the document's established format exactly, with a mandatory three-tier verification standard (Primary-verified / Secondary-credible / Unconfirmed) built directly into the prompt — a direct, structural response to the two real errors found across this session's two external-report reviews, rather than just a note to be careful.

2 new claims (CL-659, CL-660). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 128 — Claude Sonnet (Full-library tag audit: 308 instances scanned, taxonomy redesigned, 2 new fleet tasks) — July 2, 2026

Built a real extraction and categorization pipeline rather than eyeballing a grep count — scanned all 77 documents for ⚠️ markers (308 total), excluded two categories that inflate the raw number without representing real work (TRACK_003_Corrections_Log's 7 meta-mentions of the marker system itself; DERIV-012's 25 legitimate dynamic election-tracking table cells), and sampled deeply across the remaining 276 by document and by pattern.

The central finding: a single symbol was doing at least four different jobs — genuine temporary "go verify this" flags, permanent methodological caveats that will never resolve to ✅, deprecated figures correctly kept visible to prevent re-citation, and dynamic real-world data. HMK-038's 52 instances turned out to be almost entirely one coherent thing (an unverified historical per-diem table), not 52 separate problems. HMK-012's 37 were almost entirely permanent caveats ("single-model, verify before use," "superseded, do not cite"), not a task queue. This conflation is precisely why triage felt hard — the symbol alone couldn't distinguish "real work" from "important, permanent, non-actionable context."

Fixed three genuine stale tags directly (Woodhall-Melnik citations in DERIV-005 and HMK-010 marked "⚠️ corrected" when the correction was already complete and verified — changed to "✅ corrected"). Established a new tagging taxonomy in TRACK_001_Master_Issues Tier 5 as a standing convention: ✅ unchanged, ⚠️ VERIFY now strictly reserved for genuine actionable gaps, and a new 🔍 CONTEXT marker for permanent methodological notes that should never be treated as a to-do queue. Existing instances were not retroactively reclassified — that's logged as its own future work item, not done in this pass, to avoid a rushed, error-prone bulk edit.

Rebuilt Tier 6 with accurate current numbers (the prior version referenced a stale "70 documents" figure; the library has grown to 77), and extracted a clean, genuinely actionable VERIFY list — roughly 22 real items, filtered against already-tracked gaps (councillor names are Gap #14, time-gated; Toronto Social Corps is already in Strategic Observations) so nothing appears twice under two different names.

Designed two new LLM fleet tasks. Task O is narrow and mechanical: verify the HMK-038 per-diem table against primary 2010 City documents, row by row, matching the standard of the one already-confirmed row. Task N asks a model to review this library's own index and gap list and identify genuinely missing sources, with the same three-tier verification standard built into the prompt from the two external-report reviews earlier this session.

Registry, RAG export, and stats synced via the toolchain across all touched documents. Final state: 64/64 tests passing.

PASS 129 — Claude Sonnet (Narrative-gap audit: 5 stale claims fixed, 8 genuine gaps newly tracked) — July 2, 2026

Ran a scan distinct from the prior citation-tag audit — this one for self-identified incomplete-work language in prose ("not yet reflected," "does not yet capture," "has not yet been done") rather than ⚠️ markers. The seed case was the UHEI sentence the user quoted directly from HMK-040 — verified it was genuinely stale (UHEI was integrated into HMK-008 in an earlier pass; the sentence announcing the gap was never updated), then searched the same pattern library-wide.

Found 58 matches. Individually verified each rather than trusting the pattern match alone. Five were confirmed stale — the underlying work was already done elsewhere, but the original sentence never got updated: the UHEI example itself, Making the Shift/FNS (HMK-039 → HMK-010), the Black youth CJS finding (HMK-039 → HMK-001), SOHC 2014's cost ratio (HMK-039 → HMK-012), and the Ombudsman refugee-shelter-exclusion finding (HMK-017 → POLICY-001 REC M-12). Fixed all five directly, each now pointing to where the real content lives rather than claiming it doesn't exist. A sixth item was partially stale — only one of three O'Grady/Gaetz/Buccieri 2011 recommendations had actually been added (the amnesty one, REC M-15); the other two were still being described as fully open when the framing implied resolution — corrected to distinguish the closed one from the two genuinely open ones.

Cross-checked the remaining candidates against TRACK_001_Master_Issues individually rather than assuming absence — confirmed 8 were genuinely never tracked anywhere: the two real O'Grady/Gaetz/Buccieri recommendations; Bill C-12 (a real, current, March 2026 federal law with direct Toronto shelter-demand relevance, sources already found, ready for direct integration rather than fleet research); seven additional unverified HMK-038 claims beyond what Task O already covered (a 2009 system-wide average, provincial/city funding-split percentages, two staff-report quotes, a cost-savings claim, Winter Respite figures, and HSCIS capital figures); a same-year Homes First/Dixon Hall comparison idea; an unadded FOI-template candidate; POLICY-001's undone build-vs-buy evaluation; and a single-sourced drop-in-centre claim needing a cross-check. All eight added to a new TRACK_001_Master_Issues entry (Gap #69) rather than left only in this conversation's text.

Expanded Task O (TASK_001_LLM_Fleet_Tasks.md) from 9 to 16 specific figures to absorb the newly-found HMK-038 items into the same document-level task rather than splitting into a second assignment, and flagged Bill C-12 explicitly as direct integration work, not a fleet task, since its sources are already in hand.

1 new claim (CL-661) documenting the audit itself. Registry, RAG export, and stats synced via the toolchain across all touched documents. Final state: 64/64 tests passing.

PASS 130 — Claude Sonnet (Fleet prompts redesigned for tracking + open-endedness; Bill C-12 integrated) — July 2, 2026

Redesigned the standing template block in TASK_001_LLM_Fleet_Tasks.md that every fleet task inherits from. Added a required title format (# Task Name — Model Name — Date as the literal first line of any research output) so a saved file identifies itself the moment it's opened, without depending on filenames or reading into the content — directly responding to the messy filename conventions from earlier fleet rounds (qwen37plus1.txt, qwen37plus2.txt, etc.). Added a required footnote system with three explicit confidence tiers (Confirmed/Reported/Uncertain) per claim, extending the verification-tier discipline already built into Task N to every task in the document. Added an explicit design note distinguishing when prescriptive numbered-step prompts are appropriate (narrow mechanical verification, like Task O) from when they actively hurt open-ended research by anchoring a model to only what's explicitly asked. Loosened Task B (media landscape) as the clearest example of the latter — replaced a rigid five-category checklist with a stated goal and evidence standard, letting the model's own read of what's actually rich or thin in Toronto's media landscape shape the output rather than forcing uniform coverage of a predetermined template. Checked several other tasks (E, L) and confirmed they were already appropriately calibrated for their purpose — didn't loosen what didn't need it.

Caught a real, if minor, self-inflicted error immediately via the test suite: the new template's own example used a real model name ("Claude Opus 4.8") as a sample, tripping the model-leakage test. Fixed to a generic placeholder before it went further.

Moved into the highest-priority ready item flagged at the end of the prior pass: Bill C-12. Integrated into HMK-032 as a new Part 3B, placed precisely where it belongs structurally — immediately after Part 3's existing IHAP/federal-funding argument, since Bill C-12's real significance is a direct, mechanical pathway from the IHAP-funded refugee-claimant population into the entirely unfunded without-status population Part 2 of the same document already establishes. Corrected HMK-029's own now-stale "not yet integrated... should be" language accordingly, and closed the item within Gap #69's tracking rather than leaving it implicitly resolved only by the new HMK-032 content.

1 new claim (CL-662). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 131 — Claude Sonnet (Continuing Gap #69: 4 of 8 items now closed, including a genuinely surprising find) — July 2, 2026

Continued directly through Gap #69's queue rather than waiting for further instruction, per "continue through our priorities."

REC M-16 added to POLICY-001: the two remaining O'Grady/Gaetz/Buccieri 2011 recommendations (stop using ticketing as displacement, don't stop youth on homelessness status alone) — a requested Toronto Police Service Board directive, carefully scoped to respect the real jurisdictional limit that Council cannot direct police operations directly, only request Board consideration. Count propagated 49→50 across 10 files plus a new council motion clause and cost-benefit matrix row, following the exact established pattern from M-15.

HMK-016's FOI template flag resolved as stale, not genuinely open: the template it asked for already exists as DERIV-004's FOI-004, requesting the identical transition binder. Corrected to point to it directly rather than duplicate it.

DERIV-016's drop-in centre claim produced the most interesting finding of this pass. Confirmed the "59+" count exactly (Toronto Drop-In Network: 59 member organizations, 56+ centres) via direct search. But the "$15 million combined operating budget" figure didn't surface anywhere as TDIN's own number — instead, cross-referencing against this library's own already-verified content found HMK-036 already documents a $14.4M TSSS budget line for "Drop-Ins and Housing Focused Client Supports." $14.4M and "less than $15 million" are close enough that the original claim likely conflated a single City department's specific budget line with a citywide coalition's combined budget across dozens of mostly-independently-funded organizations — two very different things. Corrected in DERIV-016, and the confirmed TDIN count added to HMK-036 as a genuine scale cross-check. Worth naming as a pattern: checking a flagged claim against the library's own existing verified content, not just external search, resolved this faster and more precisely than search alone would have.

Two items remain open in Gap #69 (HMK-005's comparison proposal, POLICY-001's build-vs-buy evaluation) — both genuinely require more substantial analytical work than a single search or direct addition, appropriately left for a dedicated pass rather than rushed.

2 new claims (CL-663, CL-664). Registry, RAG export, and stats synced via the toolchain across all touched documents. Final state: 64/64 tests passing.

PASS 132 — Claude Sonnet (Task H results reviewed: 3 facts confirmed, 1 severe fabrication caught) — July 2, 2026

Received seven external research submissions for Task H (political accountability mapping), the task explicitly flagged in advance as hardest to search well. Applied the same individual-verification discipline as every external submission this session, rather than merging them.

Confirmed and integrated into DERIV-008: the May 12, 2023 homelessness emergency declaration (24-1, Holyday's third confirmed dissent, cross-verified against five independent sources including the City's own press release); Chris Moise's $10,800 Fitzrovia Real Estate campaign donations (11.8% of his 2022 total, confirmed via direct CBC investigation, with the added context that he subsequently supported a specific Fitzrovia development Council approved); and ACORN Canada's 2018 developer-donation report (34% council-wide, with Bailão's 42% and Ainslie's 21% figures matching the primary PDF exactly).

Found the most severe external-report failure of this entire session. One submission — a formatted PDF — has a completely fabricated Toronto ward-councillor map: Stephen Holyday listed as "Ward 7, Richmond-Steeles" and Frances Nunziata as "Ward 4, Allenby-Danforth," neither a real Toronto ward name, cross-checked and confirmed wrong against Wikipedia's direct account of the council term. The same document presents two former councillors (Ana Bailão, Mike Layton) and a 2017-era petition as current council business. This is a different kind of failure than Bill 28 or the Homelessness Ending Plan mischaracterization found earlier the same day — those were wrong claims about real things; this is a wrong core dataset underlying an entire report's structure. Flagged prominently in DERIV-008 with an explicit instruction not to use any of that submission's content without independent reconstruction.

Used this library's own already-verified content as a cross-check tool throughout — DERIV-008 already had a confirmed, named breakdown of the November 2025 MM34.4 vote and a partial confirmed table for EC7.7/EC13.8/PH23.3, which let several of the other submissions' specific claims (particularly around Pasternak's Wilson Ave votes and Holyday's consistent dissent pattern) be cross-validated against ground truth already in hand rather than needing fresh searches.

Updated INTAKE_001's Task H entry to reflect actual status — partial success with a specific severe caveat, not a clean completion. Designed Task P as a narrow, specific follow-up covering exactly what remains unresolved (the July 2025 vote's exact roll call, whether the December 2024 "demonstrations" vote is actually homelessness-relevant at all, the February 2023 warming-centre vote, and the still-undone PDF-by-PDF campaign finance review every submission flagged but none completed) — deliberately more open-ended in structure per this session's earlier prompt-design work, not a rigid numbered checklist.

4 new claims (CL-665 through CL-668). Registry, RAG export, and stats synced via the toolchain across all touched documents. Final state: 64/64 tests passing.

PASS 133 — Claude Sonnet (Header/footer reminder added to all 15 fleet tasks) — July 2, 2026

A structural gap: the title/footnote requirement added earlier this session lived only in the document's intro material, but "Instructions for use" explicitly tells the operator to copy each task as a self-contained block to a fresh LLM session — meaning a task copied alone would silently drop the requirement. Fixed via a script-based insertion (not 15 manual edits, to guarantee consistency and avoid the recurring header-drop bug at scale) adding a compact header reminder right after each task's title and a footer check right before its closing separator. Verified all 15 task headers and all three sequencing notes survived the bulk edit before treating it as done.

PASS 134 — Claude Sonnet (Fleet task header/footer fix; priorities review; POLICY-001 build-vs-buy resolved) — July 2, 2026

Applied the requested header/footer reminder to all 15 fleet tasks via a script (not manual per-task edits, to guarantee consistency across the batch) — every task now carries the title-format and confidence-tier footnote requirement directly in its own block, since tasks are meant to be copy-pasted individually and the requirement would otherwise only exist in material that might not travel with a single copied task. Verified all 15 headers and three sequencing notes survived the bulk edit.

While pulling current status for the priorities review, caught a stale reference: Task G's item 1 still described Bill C-12 as "urgent... not yet analyzed," when it was directly integrated into HMK-032 two passes earlier. Narrowed the task's description to what's actually still needed (Toronto-specific numbers, removal-process timeline) rather than re-describing already-completed work as open, and synced the same correction into INTAKE_001.

Closed another item from Gap #69: POLICY-001's build-vs-buy evaluation for a shelter coordination system. Research found HIFIS's documented scope is coordination within the homelessness sector — the same category SMIS already covers — with nothing suggesting native integration with hospital, corrections, or income-support systems. Cross-referenced against general health-data-sharing literature and a directly on-point academic case study (a US hospital/HMIS patient-matching integration project) confirmed the real barrier is a data-sharing-agreement and health-privacy-law problem, not a choice between SMIS, HIFIS, or a custom build. Rewrote the recommendation's own text to reflect this — not just answering the original build-vs-buy question, but reframing it, since the honest answer changes what the recommendation should actually ask for.

1 new claim (CL-669). Registry, RAG export, and stats synced via the toolchain across all touched documents. Final state: 64/64 tests passing.

PASS 135 — Claude Sonnet (Full close-out pass: reorganization, archival, and external-review prep) — July 2, 2026

A full "prep for handoff" pass covering document reorganization, archival of completed work, and readiness checks across the library's tracking and orientation documents — treated as the last pass before external review and LLM handoff, per direct instruction.

Fixed Gap #68 for real, not just re-flagged. Found and corrected all 8 residual model-name-leak instances across HMK-001 and POLICY-001 (one more than the original count — a fourth "Gemini Free Thinking" instance had been missed even in the earlier flagging pass). More importantly, diagnosed why test 7.1's regex hadn't caught these: it required a word boundary that compound model names like "Qwen37pT" and "Qwen37plusThinking" don't have, and it only matched the literal "Meta/Gemini," not standalone "Gemini." Rewrote the pattern. It immediately caught one further live instance the manual sweep alone had missed, confirming the pattern-level fix mattered, not just the instance-level ones.

Fully rebuilt TRACK_001_Master_Issues (version 8.0). The document had grown to 194 lines with entries numbered non-sequentially (6, 7, 8, 9, 10, 11, 16, 30, 33...70) and 39 items still sitting in "active" tier tables despite being marked closed in their own status text. Moved every genuinely closed item to _internal/MAIN_TRACK_005_Resolved_Issues_Archive.md in full — nothing deleted, only relocated — and rebuilt the active document from scratch with clean sequential numbering by tier. Also removed a redundant hardcoded document count that duplicated REF_001_Index.md's authoritative figure (and was itself already stale, an "84" that should have been the actual 80) — the exact kind of two-copies-that-drift problem this document already warns against.

Rebuilt INTAKE_001's fleet-task table, which had structural problems (detailed narrative crammed into the Status column for two rows instead of Notes) and was missing Tasks N, O, and P entirely. Added closure status for Tasks A and M, added the three new tasks, and added an explicit priority-order line matching the fleet document's own sequencing guidance.

Refreshed MAIN_REF_001_Index.md — updated the stale June 30 date and "last major update" summary (which only mentioned HMK-039 and didn't reflect any of this session's substantial work), and honestly re-labeled the "Critical Current-State Update" table as a captured snapshot rather than implying a fresh re-verification that didn't happen.

Added Part 9 to REVIEW_001_External_Review_Prompt.md, following the document's own established dated-addendum pattern exactly (matching the style of Parts 4, 5, 6, 7, 8 rather than inventing a new format). Flagged the three-for-three pattern of external research batches this session each containing at least one serious, confidently-stated error, and made the explicit recommendation that a reviewer's highest-value activity right now is spot-checking this library's own claims cold, not just re-checking internal consistency — since this library has gotten disciplined about catching external errors, making its own claims the least-recently-stress-tested part of the system.

Caught two more stale recommendation-count instances during a final broad sweep (REF_001_Index.md's body text, POLICY_002_Council_Resolutions_Draft's source line) that the automated test's narrow context-trigger regex had missed — one because it used "MAIN_POLICY_001_Master_Recommendations.md" instead of "POLICY-001," the other because "these policy recommendations" isn't one of the regex's specific trigger phrases. Fixed both instances directly, then narrowly hardened the test pattern to catch both specific gaps without loosening it broadly enough to reintroduce the 36-false-positive problem the pattern's own docstring already warns it once had.

Produced MAIN_TRACK_002_Priorities_2026-07-02.md as the actual deliverable — a fresh, dated snapshot superseding July 1's, per that document's own explicitly-stated design ("supersede it rather than append to it"). Left the July 1 snapshot untouched as historical record, with a pointer note added for discoverability.

Registry, RAG export, and stats synced via the toolchain across every touched document. Final state: 64/64 tests passing, citation checks clean (remaining flags confirmed as parenthetical-date false positives), full library reorganized and current.

PASS 136 — Claude Sonnet (Final readiness pass: version metadata, tag reclassification started, freshness confirmed) — July 2, 2026

The final pass before external review and fleet-task deployment, per direct instruction.

Closed the version-metadata gap completely. All 39 HMK documents now carry a Version field (6 already had one; 33 added this pass). Used a deliberately honest approach rather than fabricating false precision: a simple "1.0" baseline for lightly-touched documents, and for 15 documents with clear record of multiple substantive revision passes this session, a version number labeled explicitly as "approximate, reconstructed retroactively" rather than presented as an exact count. Verified via direct grep that all 39 documents now pass, not just spot-checked.

Started the retroactive ⚠️ → 🔍 CONTEXT tag reclassification, deliberately not rushed to completion. Reclassified 10 clear, unambiguous instances across HMK-001 (7) and HMK-012 (3) — pure disambiguation notes with no verification language. While doing this carefully, found a real trap: several ⚠️ instances that looked like permanent caveats at first read ("single-model — verify before public use") actually still invite genuine future verification, and reclassifying them would have silently removed them from any future action queue. Left those as ⚠️ VERIFY rather than force a clean-looking but wrong answer. This is exactly the kind of judgment call that shouldn't be rushed through as a bulk pattern-match, and the remaining ~100+ instances library-wide are logged honestly as still needing the same one-at-a-time care, not claimed as done.

Ran a full freshness sweep: complete test suite (64/64), a citation check across every HMK/DERIV/POLICY/BRIEF document (one file flagged 25 instances, all confirmed on inspection to be the same parenthetical-date false-positive pattern seen throughout this session, not real gaps), and a spot check of MAIN_REF_002_Sources.md (a reference compendium without a narrative date to go stale, left untouched as appropriately out of scope for this pass).

Registry, RAG export, and stats synced via the toolchain across every touched document. Final state: 64/64 tests passing, model-leakage test clean, citation checks clean of anything beyond known false positives, 178 pre-edit archive snapshots preserved across the full session. The library is current and ready for export.

PASS 137 — Claude Sonnet (Rapid pass, 8-min window) — July 2, 2026

Fast targeted fix: TRACK_001_Master_Issues' P-05 note was itself stale — the evidence it described as "not yet written into P-05" was already there. Corrected. Verified S-05's note remains accurate. 64/64 tests passing throughout, synced via toolchain.

PASS 138 — Claude Sonnet (Real fleet-task progress: Calgary comparison found directly) — July 2, 2026

Rather than only continuing tracking/cleanup work, attempted Fleet Task I directly using available search — a genuinely open, unattempted task, unlike Task K which had already come up empty twice through normal channels.

Vancouver yielded no clean per-diem figure; BC's shelter funding runs through BC Housing/HEART/HEARTH program structures that may not use per-diem framing at all, itself worth noting for whoever eventually completes this task for the remaining cities.

Calgary produced two genuinely strong, real findings. First, a cost comparison: Charity Intelligence's calculation from the Calgary Drop-In Centre's own CRA filings (F2025) gives roughly $70/person/night, against Toronto's ~$136 base figure — labeled honestly as one-operator-vs-system-average and a third-party calculation rather than a clean city-to-city match, not overstated as more comparable than it is. Second, and stronger: a Policy Options/IRPP academic source using the identical administrative shelter-use dataset across both Calgary and Toronto for the same 2011-2016 window found average shelter stays of 17 days in Calgary versus 45 in Toronto — a real, methodologically-matched cross-city benchmark that directly extends this document's own bed-turnover finding (Toronto's turnover roughly halving since 2011) with an actual comparison point rather than Toronto's trend viewed in isolation.

Integrated both into HMK-038 as a new Part 10, version bumped to 3.3, and updated Task I's status in INTAKE_001 to reflect real partial completion rather than leaving it listed as simply "assigned" when meaningful progress now exists. Ottawa and Montreal remain untried and still worth a fleet assignment specifically for those two, along with independently confirming the Calgary study's exact authorship before the citation is used in anything public-facing.

2 new claims (CL-670, CL-671). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 139 — Claude Sonnet (Task I: citation confirmed, Ottawa added; 3 of 4 target cities now attempted) — July 2, 2026

Closed the loose end flagged at the end of the previous pass: tracked down the exact peer-reviewed citation behind the Calgary/Toronto shelter-stay comparison. Confirmed as Jadidzadeh, A., and Kneebone, R. (2023), "Homeless Shelter Use by Age and Community... A Comparison of Calgary and Toronto," Canadian Public Policy 49(2):180-196 — Ron Kneebone is Scientific Director of Social Policy and Health at the University of Calgary's School of Public Policy. This moves the citation from "needs verification before public use" to fully confirmed, peer-reviewed. Also surfaced a related finding from the same research group's earlier 2018 paper specific to chronic shelter users (329.7 days Toronto vs. 116 days Calgary) — added but explicitly flagged as not yet independently confirmed against its own primary text, rather than borrowing confidence from the now-confirmed 2023 paper.

Attempted Ottawa directly. Found a real, usable figure in an official 2025 City of Ottawa council document: approximately $36,500/year to support one household in the emergency shelter system, converting to roughly $100/night. Deliberately did not present this as directly comparable to Toronto's $136 per-bed figure — it's a per-household number, and collapsing that distinction would repeat the exact "different denominator" error this library's own HMK-012 already warns against. Presented as a genuine Ottawa data point with the caveat stated plainly, not glossed over for a cleaner-looking comparison.

Task I now stands at 3 of 4 target cities attempted (Vancouver: no clean figure found, itself informative; Calgary: two strong confirmed findings; Ottawa: one real figure with an honest scope caveat). Only Montreal remains genuinely untried.

2 new claims (CL-672, CL-673). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 140 — Claude Sonnet (Task I fully completed: 4 of 4 cities attempted) — July 2, 2026

Attempted Montreal, the last untried city for Fleet Task I. Found real financial data for Mission Old Brewery via Charity Intelligence ($16.4M F2024 program spending) but it covers three combined programs, not emergency shelter specifically, with no clean nightly-occupancy figure to divide by the way Calgary's did. Deliberately did not force an estimate by guessing at a program-split assumption this document couldn't defend — the same discipline already applied to Toronto's own total-budget figures in Part 3. Recorded this as a genuine negative result with a stated reason, joining Vancouver, rather than a silent gap.

This completes Task I's original four-city scope directly rather than through the fleet: two cities (Calgary, Ottawa) produced real, usable comparative figures; two (Vancouver, Montreal) were genuinely attempted and came up empty for structural reasons specific to how those cities report shelter costs — itself a real finding about uneven reporting standards across Canadian municipalities, not a research shortfall. Marked closed in INTAKE_001 accordingly.

1 new claim (CL-674). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 141 — Claude Sonnet (Task K attempted directly: 2 of 4 remaining operators, honest negative results) — July 2, 2026

Picked up Task K's operator-profile half — the part distinct from the already-twice-failed per-diem search. Attempted John Howard Society of Toronto and Fife House directly.

Both confirmed as real organizations with real scale (CRA business numbers, founding history, service scope) across multiple search angles each. Neither produced a clean financial figure — no Charity Intelligence profile exists for either, unlike the larger operators already profiled in this document. Recorded as genuine negative results, following the document's own already-established "confirmed real, financial profile not located" pattern from an earlier pass, rather than either forcing a weak estimate or omitting the attempt entirely.

One real cross-reference surfaced along the way: Fife House already appears, unconnected, in this library's own HMK-038 historical per-diem table ($20.25/day, 2010) — now noted explicitly. Also caught a genuine uncertainty rather than resolving it artificially: three different sources gave three different signals about Fife House's current Executive Director (a December 2024 departure, an April 2025 "Acting" title, and an undated Pride Night article naming someone else) — flagged plainly as unresolved rather than picking the most recent-sounding name and presenting it as settled.

Native Child and Family Services of Toronto and Christie Ossington Neighbourhood Centre remain fully unattempted. The per-diem search itself — the harder, higher-value half of Task K — was not reattempted this pass, consistent with the standing instruction not to repeat a search that's already failed twice through the same channels without a different approach.

2 new claims (CL-675, CL-676). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 142 — Claude Sonnet (Task K operator profiles completed: real scope finding on NCFST) — July 2, 2026

Completed Task K's operator-profile half — all four named organizations now attempted (JHS Toronto and Fife House from the prior pass; Native Child and Family Services of Toronto and Christie Ossington Neighbourhood Centre this pass).

Christie Ossington Neighbourhood Centre confirmed as a direct, unambiguous shelter operator — two men's shelters, 20 transitional units, 198 hotel-shelter rooms, a documented six-pillar service model — but no financial figure surfaced despite the attempt.

Native Child and Family Services of Toronto produced the most interesting finding of the pass, and not the one initially expected. It's confirmed as Canada's largest Indigenous multi-service organization, with a real $60M-scale budget figure (sourced to a credible secondary announcement, labeled Reported rather than Confirmed accordingly) and a genuinely notable governance finding — it moved to a Leadership Council model in 2024 rather than a conventional single Executive Director. But everything found describes it as a child welfare and family-wellbeing agency, not a shelter operator. Rather than force it into HMK-005's frame because Task K named it, flagged this directly: it likely fits better cross-referenced with HMK-031 (Indigenous Self-Determination Overlay) than treated as a peer of Homes First or Dixon Hall.

Caught and fixed a real error during this pass's own toolchain run: used "REPORTED" as a claim-ledger status value, which isn't part of this library's valid status vocabulary (VERIFIED/VERIFY/UNVERIFIED-REMOVED/REMOVED) — that's fleet-task confidence-tier language, a different system. Test 1.4 caught it immediately; corrected to VERIFY before treating the pass as complete.

Task K's operator-profile half is now closed. The per-diem search — the harder, higher-value remaining half — was not reattempted, consistent with not repeating a twice-failed search through the same channels without a genuinely different approach.

2 new claims (CL-680, CL-681), one corrected mid-pass. Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 143 — Claude Sonnet (Task P: category error confirmed and corrected; a full vote resolved) — July 2, 2026

Attempted Task P (the Task H follow-up) directly. Two of its four items resolved.

The more important of the two: confirmed CC24.2 ("Policy Framework - City Response to Demonstrations," December 2024) is not homelessness or encampment-relevant. Direct retrieval of the City's own agenda item and report shows it concerns hate-motivated demonstrations targeting religious and cultural institutions — a "bubble zone" bylaw protecting places of worship, faith-based schools, and cultural institutions, plus a hostile-vehicle-mitigation grant program. This confirms a category-error risk flagged (but correctly left unresolved rather than assumed either way) during the earlier Task H review. One of the seven external submissions reviewed for Task H had built a substantial voting-bloc analysis around this exact vote as if it were a homelessness item — that analysis is now explicitly flagged as wrong on relevance, even though its vote count (17-5) was accurately transcribed. Removed from DERIV-008's "unverified" list and replaced with the corrected finding.

The February 8, 2023 warming-centre vote (15-11, rejecting 24/7 warming centres and a formal crisis declaration in favor of Councillor Thompson's alternative motion) is now fully confirmed across six independent sources, upgraded from "reported by one submission." Two bonus figures surfaced in the same sourcing — a $400,000/month per-centre 24/7 operating cost estimate and a comparative 2022 funding figure ($457M City vs. $97M provincial vs. $72M federal) presented at the same council meeting — flagged for future cross-reference against HMK-008/HMK-034 rather than assumed consistent without checking.

Two items remain: the exact July 2025 six-site shelter rezoning roll call, and the campaign-finance PDF review every Task H submission agreed was needed but none completed. Both still worth a fleet assignment.

2 new claims (CL-682, CL-683). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 144 — Claude Sonnet (Task P essentially complete; a self-caught citation error corrected in the same breath) — July 2, 2026

Resolved Task P's third item: the July 2025 PH23.3 shelter-site vote's exact tally, confirmed as 22-3 via a contemporaneous live-tweeted vote tracker (Progress Toronto), cross-referenced against the City's own full procedural record (points of order from Councillors Perruzza, Perks, and Holyday; Speaker Nunziata's rulings on grouped vs. separate recommendation votes). This resolves the earlier "19-23 vs. 22-3" ambiguity in favor of 22-3.

Caught a real error in my own writing before it left the room. While integrating a genuinely striking finding — Councillor Bradford voted against all six shelter sites — I initially described it as "consistent with this document's own already-confirmed table showing Bradford voting No on all six sites." Checking the actual table immediately after writing that sentence: Bradford isn't in it at all. Only Holyday, Morley, Pasternak, Colle, Bravo, Cheng, and Kandavel are listed for PH23.3. Corrected the claim before moving on, and — rather than just deleting the false consistency claim — actually added Bradford as a proper new row with the real, sourced data, so the correction produced something more accurate rather than just something less wrong.

Task P is now effectively complete: three of its four original items resolved directly (CC24.2's category error, the Feb 2023 vote, and PH23.3's tally), leaving only the campaign-finance PDF review — a large, mechanical task (25 councillors' individual 2022 filings) better suited to a dedicated fleet assignment than further direct search.

1 new claim (CL-684). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 145 — Claude Sonnet (Full non-core document audit: 41 documents reviewed, 9 fixed) — July 2, 2026

Reviewed every document outside the HMK/DERIV/DATA/POLICY-001 core — 27 in the main outputs directory plus 14 in _internal, 41 total — for currency, staleness, and duplication.

Fixed directly (9 documents): PROC_008_Test_Suite_Documentation.md (documentation was at "56 tests/10 categories," actual is 64/12 — added a Part 6 addendum explaining what each new category exists to catch, following the document's own established pattern, and explicitly pointed to _tools/test_suite.py as the authoritative count going forward rather than a number to keep manually synced); IDEATION_001 (a copy-paste-to-external-AI prompt template stating "216 verified claims" against the real 656+ — fixed to a durable phrasing that points to checking REF_001_Index.md rather than hardcoding a number that will drift again); REF_005_ROI_Lock_Memo (header cosmetically stale — the actual locked dollar figures were verified accurate and consistent with HMK-012's current state, a good sign the underlying discipline held even where the header didn't); PROC_006_Replication_Protocol (two "50+ document" references understating the current 80+); POLICY_002_Council_Resolutions_Draft — checked, already fully current (27 clauses, both M-15 and M-16 present); BRIEF-001 ("three outstanding FOI requests" — DERIV-004 has four, FOI-004 having been added since this was written; also named FOI-004 specifically as the most consequential rather than leaving it generic).

Marked explicitly superseded rather than rebuilt (4 documents): STOCKTAKE-002, OPUS_HANDOFF, REVIEW-001, and RESEARCH_TASKS_DEFG all described project states from an early phase (188-397 claims, 41-67 documents, 20 recommendations) now far behind reality, and all four duplicated a function now served by documents already being actively kept current this session (TRACK_002_Priorities_2026-07-02 + TRACK_001_Master_Issues v8.0 for stocktaking; REVIEW_001_External_Review_Prompt for review handoffs; TASK_001_LLM_Fleet_Tasks for task assignment). Rather than a costly rebuild duplicating work already done elsewhere, each got a clear superseded-note pointing to its current replacement, with the original content preserved unchanged below the note as genuine historical record — the same treatment already applied to PRIORITIES_2026-07-01 earlier this session.

Checked and confirmed already current, no action needed: AUDIT_001, AUDIT_002, REVIEW_004_Legal_Stress_Test_Prompt, REVIEW_002_External_Review_Results, META_001, META_002, REVIEW_008_Stocktake_001 (already self-identifies as superseded), REVIEW-002, PROC_009_Removed_Model_Synthesis_Tables, PROC_007_Research_Ops_Protocol, PROCESS_001, PROCESS_002 (confirmed STATUS: COMPLETE still accurate), PROC_005_Registry_Readme, PROC_001_Deploy_Checklist, TASK_002_LLM_Task_Verification, REF_004_First_100_Days (already correctly says 50 recommendations), REF_002_Sources, DATA_003_Resource_Registry.csv.

Registry, RAG export, and stats synced via the toolchain across every touched document. Final state: 64/64 tests passing, 205 pre-edit archive snapshots preserved across the full session.

PASS 146 — Claude Sonnet (Task J: a direct false claim corrected — a real, damning AG report found) — July 2, 2026

Attempted Task J (HART Hub delivery tracking) directly. The result is the most consequential correction of this session, not just an addition.

HMK-003 explicitly stated "No FAO or AG report has examined HART Hubs." This was wrong. A real Ontario Auditor General audit — Shelley Spence's review of the province's Opioid Strategy, released December 2024 — examined the HART Hub transition directly and found the decision to close ten supervised consumption sites was made "without proper planning, impact analysis or public consultations," that the $378M investment was decided without a needs-based assessment, and that no formal consultation occurred with Public Health Ontario, site operators, or the people who used the sites. Confirmed via direct reporting on the AG's own report, cross-checked across multiple independently-published outlets syndicating the same underlying Local Journalism Initiative story.

The single most striking finding within that report: a documented 7:1 disparity in opioid death rates between Indigenous and non-Indigenous Ontarians (11.4 vs. 1.6 per 100,000), in a report that separately found Thunder Bay's Indigenous and Northern communities were not adequately consulted before their only supervised consumption site was ordered closed. This sits alongside — and substantially strengthens — the existing library content on this file's equity dimensions.

Also resolved a smaller, previously-flagged gap: three of Toronto's four HART Hub site addresses (South Riverdale, Parkdale/Bathurst, Regent Park) are now confirmed via direct site-visit journalism, which itself surfaced a real, sourced "opened on paper, not yet operating" gap — a Toronto Board of Health chair confirming promised services (supportive housing, detox beds, primary care) were delayed to "later this year, or 2026" as of the site visits, pending a still-unfinalized provincial funding agreement. The fourth Toronto site's identity remains genuinely unresolved and is logged as such rather than assumed.

This is exactly the kind of finding this whole project's verification discipline exists to catch — not an external report's error this time, but the library's own. Caught by actually attempting the research task rather than trusting an existing "confirmed" tag from an earlier pass.

3 new claims (CL-685 through CL-687). Registry, RAG export, and stats synced via the toolchain. Final state: 64/64 tests passing.

PASS 147 — Claude Sonnet (Completing the non-core audit honestly; 4 documents genuinely archived; full rename plan designed) — July 2, 2026

Answered a direct question honestly rather than assuming prior work was complete: the previous pass's non-core document audit had checked headers on PROC_002_Method, PROC_005_Registry_Readme, PROCESS_001, PROC_007_Research_Ops_Protocol, and TRACK_005_Resolved_Issues_Archive, but not their full content. Read each in full this pass. Found and fixed three more real staleness instances: PROC_002_Method.md stated "43 checks" (actual 64); PROC_005_Registry_Readme stated "299 resources registered" (actual 450); PROC_007_Research_Ops_Protocol had a leftover footer still identifying itself as "MAIN-HMK-001 v2.3" from before this content was split out of HMK-001. PROCESS_001 confirmed genuinely current, no action needed.

Completed the archival of STOCKTAKE-002, OPUS_HANDOFF, REVIEW-001, and RESEARCH_TASKS_DEFG properly — not just marked superseded in place, but individually checked for any buried, still-actionable recommendation before being physically moved to _archive/. Found real content worth verifying: STOCKTAKE-002 named "HMK-012 has zero cross-references to HMK-033/HMK-034" as its single most important finding — checked directly and confirmed this is now resolved (3 cross-references present, likely a side effect of this session's extensive HMK-012 work rather than a targeted fix). OPUS_HANDOFF's four "pending claims" were spot-checked (HSCIS $507.6M) and confirmed already resolved in two separate documents. Every other item across all four documents was already tracked in TRACK_001_Master_Issues or already resolved. Nothing was lost by archiving; the verification itself is now logged as its own TRACK_001_Master_Issues entry (Gap #28) rather than left implicit.

Designed, but deliberately did not execute, the full FLAGSHIP → MAIN file renaming convention requested. Checked the technical dependency surface first: test_suite.py has 17 references to the FLAGSHIP prefix, several load-bearing rather than cosmetic (regex patterns matching category+number format, one literal hardcoded filename inside a specific regression test's own logic), and 476 literal cross-references exist across 41 documents' prose. Built the complete old-name → new-name mapping for every affected file (MAIN_RENAME_PLAN_2026-07-02.md) and logged it as the top priority for a dedicated next pass, explaining directly why a half-completed rename — some files renamed, some cross-references left stale — would be a worse outcome than not starting, since broken references are harder to notice than files simply awaiting their turn. Also physically archived PRIORITIES_2026-07-01.md, which the rename plan itself flagged as sitting active when it should already be in _archive/.

Registry, RAG export, and stats synced via the toolchain across every touched document. Final state: 64/64 tests passing.

PASS 148 — Claude Sonnet (Full FLAGSHIP → MAIN rename: executed, tested exhaustively, two unrelated bugs caught) — July 2, 2026

Executed the file naming convention change designed last pass. All 94 files renamed per the plan. This log entry documents the full verification process, not just the outcome, since "keep going until 100% sure" was the explicit standard.

Execution approach: built a precise old-name → new-name mapping first, verified it was complete and collision-free against the actual filesystem (zero missing, zero extras, zero duplicate targets) before touching anything. Discovered during a dry-run that simple full-filename replacement was insufficient — "MASTER_ISSUES" (bare, no prefix) appeared far more often in the library's own prose than the full "FLAGSHIP_MASTER_ISSUES.md," and a third "Document ID" style (all-hyphens, no extension) existed as well. Built all three pattern variants per file, 344 total patterns, sorted longest-first to prevent partial-match corruption, applied in one pass across every file's content before any physical renaming happened.

Real gaps found and fixed during verification, not assumed away: several Document ID lines used a shorter format than the generic pattern-generator assumed (e.g., "FLAGSHIP-INTAKE-001" rather than the full descriptive bare name) — found by extracting every actual remaining Document ID line as ground truth rather than trusting the generated patterns, then building exact fixes for each. A reference to FLAGSHIP_DATA_002 in HMK-022 was missed by the first pass entirely. _tools/README.md — a file not in the original dependency assessment — had five more hardcoded references. Five additional tooling scripts (add_claim.py, backfill_registry.py, content_linkage_audit.py, refresh_stats.py, regenerate_rag_export.py) beyond test_suite.py and close_document.py/check_citations.py had hardcoded paths that would have silently broken the entire pipeline if missed — found only by explicitly listing every script in _tools/ rather than trusting the initial scan.

Two genuine bugs, unrelated to the rename itself, surfaced by the verification process and fixed on the spot: (1) an early fix attempt inserted a comma into an unquoted CSV cell (describing an archived document), corrupting column parsing for two already-REMOVED, unverified claim rows (CL-034/CL-035, unrelated to their own content) — caught immediately by the ledger's own column-count test, fixed with comma-free phrasing. (2) test_no_bare_domain_urls() had its entire function body duplicated verbatim within itself — a pre-existing bug, not something the rename caused — meaning the true test count was 63, not the 64 the library had been citing everywhere. Removed the duplicate, then had to correct the resulting "64/64" references in INDEX.md, since the suite's own self-locking test (3.6) correctly failed once the true count changed.

Final verification, not taken on faith: an exhaustive find-based sweep across every file under outputs/ (not just the files assumed relevant) for any remaining FLAGSHIP mention. What remained was individually categorized and confirmed legitimate: historical log entries describing what was actually true at earlier points in the library's history (this log, the resolved-issues archive, the external review prompt's own dated addenda), the rename plan's own necessary description of what it replaced, a correct reference to an intentionally-unrenamed archived file, and one instance of "flagship" as a plain English adjective in a document title rather than a file reference. Zero broken cross-references remain. The full 6-step close_document.py pipeline was run to completion on three different files across two categories (HMK, DERIV) as an integration test, present_files was confirmed working with the new names, and every Document ID line in the active library was checked individually for both presence and consistent formatting (four style inconsistencies found and fixed, one missing ID added).

Final state: 63/63 tests passing (the true count, corrected from the previously-inflated 64). 94 files renamed, zero data loss, _archive/ and the legacy lowercase archive/ directory both confirmed untouched.

PASS 149 — Claude Sonnet (Post-rename review; two deep-dive prompts built on a real methodology gap) — July 2, 2026

Ran a fresh internal review after the rename to confirm it hadn't introduced any subtle content issues (it hadn't — 63/63 clean, citation checks minimal and all previously-confirmed false positives).

The real finding came from reading HMK-005's existing Homes First and Dixon Hall sections in full before designing anything new, specifically to avoid duplicating work already done. Both sections are genuinely extensive — multi-year revenue trajectories, per-shelter/per-property detail, entity structure, accountability records. But neither has a computed cost-per-bed-night figure, despite Fred Victor's profile already having exactly that ($271/night, from $61.1M funding ÷ 225,707 shelter nights delivered, benchmarked against the City's $136/night and Housing First's $17.29/night). The dollar figures exist for both operators; what's missing is dividing them by actual bed-nights delivered, which requires either operator-disclosed occupancy data or cross-referencing the City's own shelter usage dashboard by site name — neither attempted yet for either organization.

Built Task Q (Homes First) and Task R (Dixon Hall) as two parallel, tailored deep-dive prompts, each explicitly instructed to replicate the Fred Victor methodology rather than reinvent one, and each targeting the specific gaps already named in HMK-005's own "what this section still does not have" notes — multi-year continuity (only 2018/2020 exist for Homes First, only 2019/2020 for Dixon Hall), Homes First's unresolved dual CRA registration numbers (the most promising unexplored lead for the long-stalled Sunshine List search), and Dixon Hall's unidentified $14.5M construction-in-process project and unresolved $56K Sunshine List total discrepancy. Both explicitly recommended for a separate Deep-Research-enabled session given the multi-year PDF retrieval and cross-referencing involved — not tasks well-suited to a quick chat pass.

Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 150 — Claude Sonnet (Task S built and attempted directly: two real, quantified findings from one search) — July 2, 2026

Built Task S — a systematic sweep of Toronto's TMMIS advanced search tool, distinct from Task N's more general "what's missing" brainstorm. Confirmed the advanced-search URL itself doesn't render via direct fetch (a client-side Angular app), so the task explicitly documents this limitation and the working fallback (site-scoped web search against secure.toronto.ca, plus the direct agenda-item/meeting URL patterns for following up specific leads) rather than presenting an untested tool as reliable.

Attempted the task directly with one search ("encampment," 2025) before handing it off, specifically to prove the approach works. It did, immediately: two genuinely new, precisely quantified findings, both confirmed against primary sources and neither previously in the library.

First: Toronto's actual COHB allocation, year by year, straight from Mayor Chow's own letter to Council — $38M (2024) → $19.75M (2025) → $7.95M (2026), a 60% cut in one year and 79% from the 2024 peak. This resolves a gap HMK-008's own existing table had been carrying ("Toronto share not disclosed") with a real number from the most authoritative source available.

Second: a standing City procurement contract titled "Security guard services - Encampment Support," $2.9-3.3M/year, confirmed via the Chief Procurement Officer's own bid award report. This is the first genuine annual, recurring encampment-enforcement cost figure this library has found — everything previously on record was the single 2021 three-park clearance event. Integrated carefully rather than combined with that figure, since they measure different things (one incident's total cost vs. one ongoing contract's yearly value) and conflating them would overstate what's actually known.

Both findings came from a single search term on a single topic, on the first attempt — a reasonable signal that Task S's core premise (that HMK-006's 31 catalogued items undersell what TMMIS actually contains) is correct, and that the remaining 20 search terms in the task are likely to surface more.

2 new claims (CL-688, CL-689). Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 151 — Claude Sonnet (Dixon Hall Sunshine List resolved; Task S validated a second time via a real user URL batch) — July 2, 2026

Confirmed, via direct arithmetic on the exact Sunshine List entries provided, that $1,185,743 is Dixon Hall's correct 2025 total — an exact match to the earlier of two conflicting data deliveries, closing the previously-unresolved $56,000 discrepancy definitively.

Processed a large batch of Toronto.ca URLs provided directly, prioritizing depth over breadth given the volume (~90 total). The single URL flagged "important" turned out to be a citizen submission to the 2026 Budget Committee — an explicitly anti-non-profit-funding advocacy piece, not a neutral source, but one that had independently computed real per-bed cost figures for both Homes First ($51,500/2024) and Dixon Hall ($64,000-67,400/2025) using the City's own Open Data "Daily Shelter Occupancy" dataset. Integrated carefully: the specific numbers were not adopted as confirmed, given the source's explicit framing and unaudited methodology, but the underlying data source is now confirmed real and usable — directly relevant to Fleet Tasks Q and R, which were already asking for exactly this kind of calculation.

A second fetch, chosen specifically for its age (1999, the oldest URL in the batch), surfaced a genuinely valuable historical find: a Council report on the provincial-to-municipal download of the Supports for Daily Living Program, effective January 2000, listing exact per-agency allocations for all 13 funded agencies — including Homes First ($582,164) and Dixon Hall ($176,499). This extends Homes First's documented growth trajectory (previously 2017-2023) back a full quarter-century, with the explicit caveat that this is one program's allocation, not the organization's total budget for that year, so it cannot be read as directly comparable to the later revenue figures.

The remaining ~49 URLs were checked against the resource registry for duplicates before being logged — 9 were found to already be present with real supporting content (confirmed via an exact URL match, not assumed), leaving 42 genuinely new entries added as "not yet ingested" targets. Two were flagged as high-priority for a future pass based on their titles alone: a 2021 "10-Point Shelter Harm Reduction" plan and a 2021 "Homelessness Solutions Service Plan," neither yet represented anywhere in this library.

While writing up this pass, caught and fixed a real error in MASTER_ISSUES' own Gap #30 — it referenced two separate task files (MAIN_TASK_003..., MAIN_TASK_004...) that don't exist; Tasks Q and R actually live inside MAIN_TASK_001_LLM_Fleet_Tasks.md. Also caught a wrong RES-ID in this same pass's own first draft (RES-0489 instead of the correct RES-0488) before it was presented — checked against the actual registry rather than assumed correct from memory.

5 new claims. Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 152 — Claude Sonnet (Real cost-per-bed-night timeseries built for both operators, 2017-2020) — July 2, 2026

Attempted direct access to the City's Open Data "Daily Shelter Occupancy" dataset per direct instruction. Tried six distinct pathways before finding one that worked: bash_tool's direct network access to toronto.ca domains is blocked entirely (confirmed empirically, contradicting what the network configuration appeared to state); the CKAN portal's static file downloads are blocked by robots.txt; the datastore-dump API endpoint returns unparseable binary; the proper parameterized datastore_search API endpoint couldn't be reached with a custom resource ID through this session's URL-matching constraints. What worked: a community-hosted mirror of the City's own 2017-2019 data (via the rfordatascience/tidytuesday R project, itself sourced from the opendatatoronto package), fetched successfully because raw.githubusercontent.com sits on a different, unrestricted network path than toronto.ca domains.

While this was in progress, the user uploaded the 2020 file directly, closing the exact gap the CKAN barrier had created.

Processed both files with Python: 115,917 rows (2017-2019) plus 41,062 rows (2020), 156,979 real daily occupancy records total. Isolated Homes First and Dixon Hall specifically, summed actual bed-nights delivered per year, and divided already-verified revenue figures by those totals — the same methodology already validated in this document for Fred Victor.

Result: a real, primary-sourced, four-year cost-per-bed-night series for Homes First ($140.02 → $164.91 → $158.32 → $274.81, 2017-2020) and a two-year series for Dixon Hall ($159.64 → $360.60, 2019-2020). Both 2020 jumps are mechanically explained, not anomalies to explain away — occupancy fell sharply (COVID physical-distancing capacity reductions) in the same year spending rose (COVID hotel-shelter expansion), and dividing more dollars by fewer nights produces exactly this kind of increase. The earlier citizen-advocacy submission's unverified 2024/2025 figures were kept as a separate, clearly-labeled cross-reference rather than blended into the confirmed table — different confidence tiers, kept visibly distinct.

Saved both source datasets as proper companion files (MAIN_DATA_004, MAIN_DATA_005) rather than leaving the underlying data undocumented behind the derived figures. Updated Tasks Q and R to reflect that their core ask is now substantially met for 2017-2020, narrowing their remaining scope to 2021-2026 specifically and naming the exact technical barrier and two concrete ways around it for whoever continues this with better tooling access.

~20 years was not achievable — the dataset itself only originates in 2017, a hard constraint of what the City has ever published in this form, not a research shortfall. Flagged clearly rather than implied otherwise anywhere in this pass's writing.

4 new claims. Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 153 — Claude Sonnet (Bed-nights timeseries complete 2017-2026; a real scope-change trap caught; mortality series extended) — July 2, 2026

The user supplied eight files directly, closing the exact gap the prior pass had named as its remaining blocker: 2021, 2022, 2023, 2024, and 2025 full-year occupancy files, a partial 2026 file (January 1 - July 1, effectively current), and two official mortality datasets.

Before computing anything, checked the new files' schema against the one already processed — a necessary step that caught a real trap. The 2021+ Open Data product ("Daily Shelter & Overnight Service Occupancy & Capacity") has a broader scope than 2017-2020's ("Daily Shelter Occupancy") — it explicitly folds in 24-hour respites, COVID hotel programs, and warming centres. A first, naive computation showed Homes First's bed-nights nearly tripling in 2021 (142,016 → 412,221) — not real growth, a scope artifact. Caught by breaking the total down by program area before accepting the number, which showed base shelter alone at 152,180 — a plausible 7% rise consistent with the rest of the series. Isolated "Base Shelter and Overnight Services System" specifically for the comparable series, while keeping the broader total as its own separate, real finding: by 2024, over two-thirds of Homes First's overnight-service-nights were delivered outside base shelter entirely, a genuine and significant fact about how much the organization's actual footprint has shifted, not something to discard just because it complicated the comparable series.

Dixon Hall's version of the same breakdown showed the reverse pattern — base-shelter nights rising every year 2021-2025 while the hotel/respite share fell from 71% to 22% — found only because both operators were computed the same way and compared, not assumed to follow the same trajectory.

Cost-per-night was computed wherever financial data already existed in the library: a clean calendar-year match for Homes First 2023 ($104.53/night on a government-funding basis, using the broader total denominator since the $60.8M figure represents the whole organization, not base shelter alone), and an approximate FY2024 figure for Dixon Hall, explicitly flagged as approximate given the fiscal-year offset against calendar-year occupancy data. Both figures come with an explicit warning not to read them as a continuation of the 2017-2020 series without the scope caveat attached, since the underlying denominators measure different things.

The mortality data resolved a real, previously-flagged uncertainty: 2023's shelter-resident death count (91) had been marked "derived, not independently confirmed" — the new primary source confirms it exactly. Also added: the full 2007-2019 year-by-year breakdown, previously known only as part of an aggregate total; 2025 and a partial 2026; and average age of death by year, a genuinely striking finding (every year 2007-2026 averages 47-58 years of age at death) that wasn't in the document before.

Saved all eight datasets as proper companion files (MAIN_DATA_006 through MAIN_DATA_013) rather than leaving the source data undocumented behind the derived figures. Updated Tasks Q and R to reflect that the bed-nights side of their core ask is now fully solved through mid-2026 — their remaining scope is exclusively financial-statement retrieval for specific named years, no longer a data-access problem.

5 new claims. Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 154 — Claude Sonnet (Overdose data integrated with field-observation caveat; Tasks Q/R completed directly) — July 2, 2026

Processed one more uploaded dataset (suspected opioid overdoses by shelter location, quarterly, 2018-2025) and isolated Homes First and Dixon Hall specifically, handling the "<5" small-cell suppression convention honestly (reported as a separate suppressed-quarter count rather than estimated or ignored). Both operators show 2021 as a sharp peak (426 and 160 confirmed incidents respectively) — an independent, third dataset now agreeing with the mortality series' own 2021 peak.

Recorded the campaign director's own field observation — that reported overdose figures likely undercount true incidence substantially — as an explicitly attributed caveat alongside the confirmed counts, not as an independently-verified statistic and not dismissed either. Framed with a concrete instruction for how the library should use these numbers going forward: as a documented floor, not a ceiling.

Then completed Tasks Q and R directly, per instruction, rather than leaving them for a future fleet assignment. Found Charity Intelligence's own multi-year financial review tables for both organizations and fetched them directly — each contained precisely the missing years as a side-by-side table, more efficient than pulling individual annual report PDFs one at a time. This closed Homes First's 2021-2022 gap exactly (government funding $54.275M and $55.048M respectively, both previously unconfirmed in this library) and Dixon Hall's FY2022-2023 gap ($21.648M and $24.640M). Combined with the bed-nights data from the prior two passes, this gives a complete, real cost-per-night series for Homes First across all seven years 2017-2023, and three fiscal years for Dixon Hall (FY2022-2024) showing a real, sustained rise ($115.73 → $181.00 → $258.72/night).

What's left is now small and specific rather than open-ended: Homes First 2024-2025 (Charity Intelligence's page hadn't been updated past 2023 as of this pass; the organization's own PDF annual report exceeded this session's fetch size limit); Dixon Hall FY2017-2018 and FY2025-2026 (older statement PDFs were confirmed to exist and be fetchable, just not yet pulled); and a proper fiscal-year-aligned occupancy recalculation for Dixon Hall, which would sharpen every already-computed figure without needing any new source. Updated Tasks Q and R, MASTER_ISSUES, and the results tracker to reflect this precisely, rather than leaving stale "ready to send" framing on tasks whose core work is now done.

4 new claims. Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 155 — Claude Sonnet (HMK-041 built: a genuine ground-up rebuild, not a relocation) — July 2, 2026

Rewrote Homes First and Dixon Hall as a new, standalone document (MAIN_HMK_041_Two_Key_Shelter_Operator_Case_Studies.md) per direct instruction, doing real additional research first rather than just moving existing HMK-005 content into a new file.

New material found and integrated that did not exist anywhere in this library before this pass: full leadership and governance detail for both organizations, including named board members with their professional backgrounds; confirmed union representation for both (Homes First: OPSEU; Dixon Hall: CUPE Local 2497, a new four-year agreement ratified April 2025); and a documented 2020 labour dispute in which OPSEU's president publicly named Homes First's Executive Director directly over an alleged COVID-19 information-withholding issue — reported as the union's own characterization, with Homes First's side of that dispute noted as not yet located.

The most consequential find: identifying Dixon Hall's previously-unnamed $14.5M "construction in process" balance-sheet line, flagged in an earlier pass as a real gap this library didn't have an answer for. It is the Parliament Street Rooming House Project — and tracing its documented cost history revealed a genuine, live accountability story: an original 2020 estimate of $6.44M has grown to $14.97M as of Q3 2025, more than doubling, with a city councillor having formally requested a full accounting from the Housing Secretariat, including why the escalation wasn't reported to Council earlier. This was not resolved in this pass — the councillor's questions remain open, and a genuine discrepancy was found and flagged rather than smoothed over: the City's own council record states the opening was delayed to Q1 2026, while Dixon Hall's own press release celebrates the project as opened on November 4, 2025. Both are cited as found, with the contradiction stated plainly rather than silently resolved in favour of whichever source was more convenient.

The new document also consolidates every piece of primary-sourced financial and occupancy work built across the preceding several passes — the full seven-year Homes First cost-per-night series (2017-2023) and Dixon Hall's three-fiscal-year series (FY2022-2024) — into one coherent, complete case-study structure, with a direct side-by-side comparison section surfacing a real, unexplained divergence: Homes First's cost-per-night stabilizes after 2020 while Dixon Hall's keeps rising with no plateau visible yet. Rather than speculate on why, this was logged as an explicit open analytical question for future work.

HMK-005's Operator 1 and Operator 4 entries were replaced with short summaries pointing to the new document, avoiding duplicate content living in two places — a surgical line-range replacement, verified against exact line boundaries before executing, not a find-and-replace on prose that could have matched unintended text elsewhere in the document.

Thirteen specific, numbered follow-up tasks were written directly into the new document's own structure, ready for direct fleet-task assignment rather than left as prose suggestions requiring translation into an actionable format.

8 new claims across this pass. Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 156 — Claude Sonnet (HMK-005 duplication resolved; a personal investigations guide built and kept outside the library; neutral research tasks logged) — July 2, 2026

Confirmed the user's direct observation was correct: two substantial sections in HMK-005 ("Comparative Reserves and Disclosure Practices," "Are These Financials Fully Exposed and Analyzed") duplicated analysis now properly housed in HMK-041. Before trimming them, checked HMK-041 against the original content and found two genuine gaps — Dixon Hall's CEO succession history (Hetherington → Watson → Mina Mawani, with compensation trajectory) and a Ci star-rating stability caveat for Homes First — neither of which had made it into the new document. Added both to HMK-041, then replaced the two duplicated HMK-005 sections with a single short pointer. Confirmed via direct grep that the remaining Homes First/Dixon Hall content in HMK-005 (the 4-operator comparison table, the cross-organization Sunshine List trend table) is genuinely multi-operator and correctly stays there.

Built a personal investigations guide per direct request — deliberately kept entirely outside the campaign library's tracked structure, in a new personal/ subfolder confirmed to sit outside the test suite's scan path, not registry-tracked, not claim-ledger-tracked, not source-checked for public use. The guide itself required real research to be genuinely useful rather than generic: confirmed Ontario's hospital/long-term-care sector staffing-agency markup problem is real, current, and government-acknowledged ($9.2B paid to agencies over a decade, new markup-disclosure legislation now in progress) but doesn't yet extend to the shelter sector; found a directly relevant recent Ontario legal precedent (November 2024, Ontario Superior Court, Anti-SLAPP) protecting an employee specifically for reporting "misuse of government funds" to CRA; and confirmed a real GTA criminal precedent (the Team Lease conviction) matching the exact staffing-agency-diversion mechanism under discussion. The guide presents four possible explanations for an unexplained expense gap — legitimate cost allocation, negligent oversight, undisclosed-but-legal agency markup, and fraud — without asserting which applies, and names the single fastest check (a staffing-agency-invoice-to-shift-schedule comparison) that would narrow between them.

While building this, the personal file initially triggered two real test failures (a real-name-leakage check, a missing-Document-ID check) by sitting in a directory the test suite scans as "public." This was the test suite working correctly, not a bug to route around — the file was moved to the properly-separated personal/ location rather than adding an exemption for content that was never meant to be library-scanned in the first place.

Logged four additional research tasks in MASTER_ISSUES, deliberately framed as neutral financial-transparency asks consistent with this library's existing, already-tracked gaps (per-diem/sole-source contracts, Sunshine List cross-referencing) rather than as allegations — per direct instruction not to write suspicion into the library's own notes.

Registry, RAG export, and stats synced via the toolchain for all library files touched. The personal guide is intentionally excluded from all of this. Final state: 63/63 tests passing.

PASS 157 — Claude Sonnet (Dixon Hall extended to 7 fiscal years; a real precision error found and corrected; both profiles restructured; every citation audited for clickable links) — July 2, 2026

Fetched Dixon Hall's FY2019, FY2021, and FY2022 audited statements directly (previously only FY2020, FY2023-via-Ci, and FY2024 existed in this library), each with an exact City/Province/Federal government-funding breakdown. This closed the FY2018 and FY2021 gaps entirely and gave seven consecutive fiscal years, six of them exact rather than estimated.

The more consequential piece of this pass was a direct response to the user's question about what "total basis" versus "base-shelter basis" actually means. Explaining it precisely required confronting a real flaw in the existing methodology: the base-shelter-basis figure divides an operator's entire government funding by only the base-shelter subset of bed-nights, which isn't a true isolated cost — it's a hypothetical upper bound. That's now stated explicitly in the document rather than left for a reader to infer from the numbers alone.

Separately, and more consequentially: Dixon Hall's fiscal year runs April–March, and the previous pass's cost-per-night figures had used calendar-year occupancy as an approximation. Using the library's own raw daily occupancy data (already in hand from prior passes), computed true fiscal-year-aligned bed-night sums for every year FY2018–FY2024. The corrected figures are not just more precise — they change the actual conclusion. The prior pass had characterized Dixon Hall's trajectory as "rising every year with no plateau." That doesn't survive the correction: the true series peaks at FY2021 ($313.23/night) and falls before partially recovering by FY2024, the same COVID-spike-then-partial-recovery shape already found at Homes First, offset by roughly one fiscal year. The document's own Part Three comparison section was rewritten to state this correction directly, by name, rather than quietly overwrite the earlier claim.

Restructured both operator profiles per direct instruction to emphasize the cost-per-bed-night figure early — each Part now opens with a "HEADLINE FIGURE" subsection (§1.3, §2.2) immediately after leadership and governance, with the full methodology and detailed tables retained further down for anyone who wants the complete picture.

Audited every citation in the document for clickable markdown formatting, per explicit instruction. Found and fixed two real gaps this produced: the Appendix's source list used bare domain names throughout rather than links, and three more bare domain mentions were found in Part Four's follow-up-task descriptions — all now proper clickable links. A final programmatic scan confirmed zero bare URLs remain anywhere in the document.

While updating Part Four, found that two of its own listed follow-up tasks (the Dixon Hall FY2021/2022 gap, and the fiscal-year-aligned recalculation) had already been completed by this same pass's other work — updated their entries to reflect that directly rather than leave a stale to-do list sitting next to the work that closed it.

5 new claims across this pass. Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 158 — Claude Sonnet (Priorities review: 7 stale items caught mid-check, MASTER_ISSUES reorganized to v9.0, priorities document fully rewritten) — July 2, 2026

Before writing a fresh priorities document, checked whether the existing gap list actually reflected the current state — it didn't. Three items (Dixon Hall's Sunshine List discrepancy, the Homes First/Dixon Hall financial coverage gaps, the same-year comparison) were still marked open despite being resolved by this session's own recent work; found by directly re-reading each one against what HMK-041 now actually contains, not assumed current because they'd been true recently.

Beyond those three, found four more items already explicitly marked resolved in their own text but still sitting in the active table, plus one item (the original Tasks Q/R entry) that had gone stale in a subtler way — its own cited figures had been superseded by more precise ones found in a later pass, making it a duplicate of two newer, more accurate entries rather than a distinct open item. Consolidated rather than left three overlapping versions of the same fact sitting in the tracker.

Archived all 7 resolved items to the full-text closed-issues record, removed the 1 redundant entry, and renumbered Tier 1 cleanly 1-28 — the same discipline as the earlier v8.0 reorganization, run again because the same pattern (closed items accumulating in the active list, numbering drifting out of sequence as items get appended) recurs naturally over a long session and needs periodic correction rather than a one-time fix.

Rewrote the priorities document in full rather than patch it, given how much had changed since it was last current: the file rename, the Task S validation, the full HMK-041 build and its precision correction, and the personal guide were all missing from what existed before this pass.

Registry, RAG export, and stats synced via the toolchain for all files touched. Final state: 63/63 tests passing.

PASS 159 — Claude Sonnet (Tier 2 worked through to completion — including catching a stale claim in my own just-written priorities document) — July 2, 2026

Worked through all four Tier 2 items from the freshly-rewritten priorities document, in order.

Costing stragglers. S-02 had already made the right call — real cost differentials found and quantified without forcing them into an estimate the evidence doesn't support. Applied that same discipline to M-11, S-01, and S-05 rather than inventing numbers to look complete. M-11 now has a ready-to-use formula (headcount × wage gap × benefits loading) with the missing input named precisely rather than guessed. S-01's exact mechanism was confirmed directly from Ontario's own OW/ODSP incarceration policy directives — benefits are suspended, not cancelled, during custody, a primary-source confirmation this recommendation didn't have before — with the one missing input (jail-population benefit-recipient share) named as a specific FOI target rather than extrapolated from the much-lower general-population rate. S-05 was reclassified from an uncosted capital line to what the earlier build-vs-buy analysis had already shown it actually is: a legal negotiation cost, not a dollar figure this library has grounds to produce.

Retroactive tag reclassification. Sampled the five largest concentrations of the remaining ~200 ⚠️ tags — HMK-038, HMK-012, HMK-001, HMK-019, and both newest documents. Every single instance checked was genuinely, correctly tagged: single-model estimates, explicit "not yet costed" markers, small-sample caveats, one advocacy-source hedge. This is a real, useful finding in its own right — the task isn't a 200-item backlog, it's confirmation that the tagging discipline has been holding. What the check did surface: two real stale cross-references in HMK-019, where a summary table still said "NOT YET COSTED" for M-11 and S-01 after this same pass's own costing work had moved past that status — fixed immediately, a direct consequence of doing the costing work first and checking downstream consistency second.

The citation backlog. This one turned into the most important catch of the pass, for an uncomfortable reason: the priorities document written two exchanges ago — by this same process, this same session — stated "~165 citations across ~26 documents remain," when PROCESS-002 had already said "STATUS: COMPLETE" before that priorities document was even written. Independently re-ran check_citations.py against the full library to confirm rather than assume either claim: zero real gaps found. The stale figure was corrected in both MASTER_ISSUES and the priorities document itself. Worth stating plainly: this wasn't an old error being caught late. It was a fresh error, in output produced this same session, caught by applying the same verify-before-repeating standard to this project's own immediately-prior work that's been applied to every external source all along.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 63/63 tests passing.

PASS 160 — Claude Sonnet (Tier 1 Research Gaps: real progress on 3 of 29 items, honest about the rest) — July 2, 2026

Started a pass through Tier 1's 29 items. Worth being direct about scope: this pass advanced three items substantively rather than attempting all 29 shallowly. Depth over a checklist.

Item 11 (Sunshine List absence). Found a working independent archival tool this session's Homes First research hadn't used before. It confirms something more specific than "absent": Homes First has exactly 4 Sunshine List records, all from 2007 and 2009-2011, then nothing through 2025 — meaning the organization had ordinary list presence in the late 2000s and disappeared just before the growth documented extensively elsewhere in this library (revenue up roughly 5x, 2017-2023). That's a sharper, more useful finding than "never appears," and it's now in HMK-041 with its source. Covenant House didn't surface on the same tool via search — noted as still open rather than pursued further with diminishing returns.

Item 23 (two flagged, unread documents). Both read in full rather than summarized from search snippets. The Homelessness Solutions Service Plan turned out to be a genuine foundational document — Council's own 2021 adoption motion directed a specific costing exercise (a one-year rental subsidy for every shelter resident) this library didn't know existed, worth a targeted follow-up on whether that report was ever produced. The harm reduction plan turned out to be more than background reading: it's the actual source document behind the "2021 Harm Reduction Directive" already referenced elsewhere in this library without its origin identified, and it contains two direct, undocumented precedents — paid lived-experience staff positions (predating and supporting REC S-06) and explicit Managed Alcohol Program referral language (supporting REC M-13). Both fully integrated into HMK-004, not just logged as read.

What this pass did not get to: the remaining 26 items, spanning genuinely blocked infrastructure gaps (ward-level CSV, POS registry), fleet-task-assigned items already covered by existing task definitions, low-priority monitoring items with no action available yet, and several session-log entries that don't need further work. A full pass through all 29 remains open — this one prioritized depth on the most tractable, highest-yield items over shallow motion on all of them.

3 new claims. Registry, RAG export, and stats synced via the toolchain across all files touched. Final state: 63/63 tests passing.

PASS 161 — Claude Sonnet (Large user data delivery processed: ward geospatial analysis, Sunshine List closure, and a major security-contractor finding that corrected this library's own prior conclusion) — July 2, 2026

Processed a large, multi-part delivery in response to the prior pass's specific file requests.

Geospatial files. Before using the uploaded shelter shapefile, checked its actual date rather than assuming it was current — it was a 2012 census. Rather than discard it or quietly treat it as current, cross-matched its real coordinates against this library's own already-held 2025 shelter address list (112 unique current addresses, extracted from the occupancy data already in hand), getting a genuine, technically valid spatial join for 40 of them (36%) against the current 25-ward boundaries. Wrote up both the real result and its real limit — twelve wards showing zero placed shelters in the table, flagged explicitly as a coverage artifact, not a finding, since claiming otherwise from partial data would be a fabricated-precision error of exactly the kind this library exists to avoid.

Sunshine List. The user's direct search of the primary CSV closed Covenant House definitively (two independent methods now agree: zero records) and, unprompted, cross-validated Dixon Hall's existing figures to the cent. Homes First's status was left genuinely ambiguous by the delivery — the message didn't specify whether it was searched and returned nothing, or wasn't searched — and this was flagged as unresolved rather than assumed closed alongside Covenant House, since those are different findings.

The council URL batch and "see also" links produced the pass's most significant finding, and it corrected this library's own prior work. A Global News investigation named the security contractor behind a previously-anonymous $2.9-3.3M/year figure and revealed it sits inside a much larger $109M citywide private-security portfolio. Following that thread turned up a real, filed class action against the same category of contractor (One Community Solutions) alleging systemic wage theft against shelter/encampment security staff, with the City allegedly aware for years while continuing to fund the company. This directly contradicted HMK-027's own prior conclusion that no such case existed after three search passes. The correction was written in explicitly, not just appended alongside the old text — the document now says the earlier conclusion was wrong and explains why the search angle missed it, rather than let two contradictory claims sit in the same file.

A Toronto Shelter Network toolkit, found via the same batch, independently confirmed that Toronto shelters commonly rely on external staffing agencies for relief coverage — directly relevant to, but carefully not overstating, the standing question about staffing-agency practices in this sector.

An AU8.3 council item, checked as part of triaging the batch, turned out to already be comprehensively documented in HMK-026 — a useful independent confirmation that existing work was accurate, not a new gap.

A real mistake made and caught within this same pass: newly-assigned resource registry IDs collided with IDs a backfill process had assigned from this session's own claim additions, and one URL was accidentally logged twice. Caught by running the test suite rather than assuming the registry additions were clean, and fixed before moving on.

Eleven items from the delivery (mostly council agenda items) were logged to the registry but not deep-dived this pass, given the volume already processed — flagged accurately as not-yet-ingested rather than silently dropped.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 63/63 tests passing.

PASS 162 — Claude Sonnet (Continued council URL batch: three procurement patterns found, and a real correction to REC M-12's "nothing was done" claim) — July 2, 2026

Continued through the remaining council agenda items from the user's link batch.

Two GG (General Government) committee items surfaced real, named, on-topic procurement examples: a numbered catering company and a nursing-services contractor, both with non-competitive contracts extended past the standard 5-year threshold. A third item found the same pattern with an IT infrastructure vendor, whose contract value increased sixfold under a genuine vendor-lock-in justification. None of the three is improper on its face — each has a stated operational reason and went through the City's own required approval process — but three independent examples of the same threshold-exceeding pattern, across three different service categories, is a real finding worth holding together rather than as isolated events, and now sits in HMK-027 as such.

The most significant find of this continuation required going back into POLICY-001 and correcting a claim already sitting there. REC M-12 stated plainly that Council received the Ombudsman's refugee-shelter report "without discussion and without adopting a single one of the 14 recommendations" — true of the December 2024 vote specifically, but the same claim, left unqualified, implied nothing further happened. A council item from the same URL batch showed otherwise: a 24-part motion adopted three months later, with real deadlines (Anti-Black Racism Analysis Tool training, a Housing Charter alignment framework, a formal legal-compliance review). The correction was written carefully to avoid overcorrecting in the other direction — this is confirmed as a substantive follow-up, not confirmed as a clean implementation of the original 14 numbered recommendations, and the single most important next check (whether the framework's own Q4 2025 reporting deadline was met) is now named as the open item, not buried.

A separate item revealed that RHI's federal successor program funds at roughly 15% of RHI's own annual national rate — a real, quantifiable funding cliff distinct from the Toronto-specific under-funding already documented for RHI itself, now added to HMK-008.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 63/63 tests passing.

PASS 163 — Claude Sonnet (Final URL batch: encampment fire mortality data and the strongest procurement registry source found this session) — July 2, 2026

Finished the remaining items from the user's council URL batch.

A 2020 emergency-housing council item, checked mainly for funding-request context, surfaced a genuinely new mortality category this library had no figure for at all: 7 deaths from encampment fires since 2010, with 216 fires in 2020 alone — a 218% increase over 2019. Confirmed via direct search that this wasn't already documented anywhere in the library before adding it. This sits alongside, but is explicitly distinct from, the shelter-resident mortality series already built — a different population, a different cause of death, not to be merged into the existing series.

The single most valuable primary source of this continuation was a 2024 procurement staff report, itemizing nine named vendor contracts with exact dollar values across catering, lodging, security, IT, construction, and interpretation. It confirmed a third, independently-named security contractor (A.S.P. Incorporated) distinct from the two already found this session (Garda Canada, One Community Solutions) — meaning this library now holds three separate named security relationships for the shelter/encampment system rather than one anonymous figure, with an open, specific question about whether their coverage overlaps or is sequential. The same report also confirmed the numbered catering company found earlier in this session (2790584 Ontario Inc) has held its contract since 2021, with a competitive replacement process explicitly planned — closing the loop on where that company's contract actually stood.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 63/63 tests passing.

PASS 164 — Claude Sonnet (Service directory processed: a real correction to this library's own prior characterization of HMK-041's largest-revenue site, plus five newly-identified shelter organizations) — July 2, 2026

Processed the pasted torontoservicedirectory.ca content. Cross-checked its listed addresses against the 2012 shapefile first rather than assuming none would match — most did, cross-validating existing ward placements, but a few genuinely new addresses surfaced.

The highest-value one: 545 Lake Shore Blvd W, confirmed as the exact site behind HMK-041's "Lakeshore" — Homes First's single largest-revenue shelter site in this library's entire dataset. Checking it directly produced a real correction to this library's own earlier work: HMK-041 had characterized Lakeshore as a "new COVID-era hotel site." It opened April 2019, about a year before COVID reached Toronto — a real factual error, not a rounding issue, and it was corrected in place with the reasoning shown rather than quietly edited. The same search surfaced current, live news this library didn't have: the City has confirmed it will not renew this lease past September 2026, and the site closes to new admissions May 1 — meaning Homes First's largest-revenue site is closing within the current research cycle, not a hypothetical future event. Precise lease-cost figures came with it ($2.75M-$2.85M/year), giving a rare clean per-bed lease-cost figure isolated from operating costs.

Five more addresses from the same list turned out to be organizations or sites not previously in this library at all — Junction Place Shelter for Men, Fatima House, a second Christie Ossington site, an entirely new organization (Cornerstone Place), and a YWCA Davenport site. None were geocoded this pass; all five are now logged by name and address as specific targets rather than left as an vague "more sites exist somewhere" note.

Ward coverage moved from 40/112 (36%) to 41/112 (37%) placed sites, confirming that targeted manual lookups, not just the automated shapefile match, can meaningfully extend this specific analysis.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 63/63 tests passing.

PASS 165 — Claude Sonnet (Ward inventory closed definitively — the City's own official export replaces this session's own approximation) — July 2, 2026

The user's uploaded CSV turned out to be a direct export from the City's own official shelter locator tool — ward assignment and coordinates already provided by the City itself, not something this library needed to infer, spatially join, or approximate. This fully closes what had been a genuinely long-standing Tier 1 item, moved to the archived-resolved list rather than left open with diminishing partial-progress notes.

The real, current, quantified finding it produced: seven of Toronto's 25 wards have zero base emergency shelter sites, while Toronto Centre alone holds nearly a quarter of the entire system, including the single largest site by a wide margin. This is now a fact this library can state precisely, sourced to the City's own published locator data, rather than an impression built from partial coverage.

Worth being direct about the relationship between this and the prior two passes' spatial-join work: that work wasn't wasted, but it's now superseded for the specific question it was trying to answer (where do current base shelters sit, ward by ward). It remains the right tool for a different, narrower question — placing the broader occupancy-derived list of temporary and hotel sites, which this new dataset doesn't cover. The document was restructured to make that distinction explicit rather than let two ward-count claims of different scope sit in the same file without a clear line between them.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 63/63 tests passing.

PASS 166 — Claude Sonnet (Untried TMMIS search terms: micro-shelters confirmed already covered, a new health-homelessness framework found) — July 2, 2026

Ran a few TMMIS search terms not yet tried this session. The micro-shelters lead turned out to already be well-documented in HMK-004 — a useful negative result, confirming existing coverage rather than a gap. A genuinely new one followed close behind: a formal "Homelessness Health Services Framework" with its own steering committee and a named Health and Homelessness Working Table, unconnected to anything previously in this library, now added to HMK-004. It produced a third independent precedent for REC S-06's paid peer-navigator model, plus two live open items (a Q1 2026 status report, an unconfirmed intergovernmental group).

Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 167 — Claude Sonnet (Sunshine List closed for both flagship operators) — July 2, 2026

User directly confirmed Homes First has zero Sunshine List disclosure, closing the one ambiguity left from the prior delivery. Both Covenant House and Homes First now confirmed absent via direct primary-source checks, matching the independent archival-tool findings on record. Moved to resolved archive, Tier 1 renumbered to 28 items.

Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 168 — Claude Sonnet (Continued TMMIS sweep while user works on download issues) — July 2, 2026

More search terms tried; mostly re-confirmed already-documented history (HousingTO, the 24-month plan, RHI). One genuine, quantifiable find: a 2023 Council ask for $60M/year enabling 2,500 new supportive housing units, compared against the actual 2025 provincial response (730 units, lower initial dollar figure) — a real ask-vs-delivered gap, not previously stated as such, now added to HMK-008 alongside the other funding-cliff patterns already documented there.

Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 169 — Claude Sonnet (Extended autonomous work session while user resolves download issues — several real finds, one genuine correction) — July 2, 2026

Worked through several more angles without needing further input.

Refugee shelter capacity confirmed in real decline (6,490 peak nightly to 3,420-3,450), with a concrete City target (a scaled, permanent 2,000-bed system) — added to HMK-032, alongside a precise operational detail on the COHB constraint (only 40 more households can move into housing before March 2026, a specific number behind the dollar-level cut already documented).

Item 9 (disposable-vs-washable shelter procurement) was tried again with a genuinely different search angle and came back empty a third time — updated to state plainly that this is very likely not indexed by web search at all, rather than leave it looking like an under-tried gap that just needs one more attempt.

The most consequential find required going back into HMK-015 and correcting a citation this library had been carrying for a while. The "31% Toronto turnover rent increase" figure was accurate to its source — but that source was CMHC's 2023 report, describing a tight-market moment, not a standing fact about what rent decontrol does. Current 2025 CMHC data shows the opposite pattern in the same market: vacancy at its highest since the pandemic, turnover rents falling across every unit size. Both figures are true of their own year; only one was being presented as current. The correction states this directly — a market reversal, not a smaller version of the same trend — rather than let a stale number keep functioning as though nothing had changed since 2023.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 63/63 tests passing.

PASS 170 — Claude Sonnet (Computed the first real dollar figure for the "new jobs" tax revenue channel) — July 2, 2026

REC S-06's 200 Peer Navigator positions were already costed in wages; the tax-revenue side of that same headcount had never been calculated. Pulled verified 2025 federal and Ontario tax brackets, showed the full calculation rather than asserting a rate, and got $1,602,611/year in combined income tax — split federal and provincial. Deliberately left CPP and EI out of the total, since both carry a future benefit obligation and counting them as clean new revenue would overstate the case, the same standard this library already holds itself to elsewhere in this section.

This closes one piece, not the whole question — REC M-11 and HSCIS construction employment still have no headcount to apply the same math to. Said so directly in both HMK-029 and MASTER_ISSUES rather than let a partial answer read as a complete one.

Registry, RAG export, and stats synced via the toolchain. Final state: 63/63 tests passing.

PASS 171 — Claude Sonnet (Full library stocktake and cleanup) — July 2, 2026

Full structural audit rather than incremental work. Verified: 83 MD + 16 CSV = 99 documents (confirmed accurate against REF_001's own count, resolving what first looked like a discrepancy in my own quick count); zero broken FLAGSHIP references (all remaining mentions individually re-confirmed legitimate — historical log entries or "flagship" as plain English, consistent with Pass 148's original verification); every active document carries a Document ID; the personal/ folder remains correctly isolated from the test suite's scan path; archive folder at a manageable 12MB/289 files.

Found one real internal inconsistency: REF_001_Index.md stated "80 MAIN documents" in its summary line while its own "Library state" line further down correctly said 99 — the two had drifted apart. Fixed both to agree.

Added three new regression tests protecting this extended session's most fragile corrections (Lakeshore's true opening date, the stale 2023 CMHC rent figure, REC M-12's Council-response framing) — the same discipline already applied to earlier corrections throughout this library. Two of the three initial patterns had real bugs: one was too loose and flagged a correct, unrelated mention of a different site (Delta) as if it needed the same correction; the other's hedge-detection window was too narrow to find the correction text that was genuinely present, just further away than checked. Both fixed by testing against the actual document text rather than assuming the pattern was right because it looked right.

Registry, RAG export, and stats synced via the toolchain. Final state: 66/66 tests passing.

PASS 172 — Claude Sonnet (A real error found and fixed in this library's own HSCIS capital figure, with two bugs of my own caught along the way) — July 2, 2026

Attempted one of Task O's seven items directly. HMK-038 carried a "$957 million HSCIS capital, $690 million debt-financed" figure, already flagged internally as unreconciled against a separately-cited $674.5M elsewhere in the library. Checked directly against four independent primary sources spanning three budget years — all four agree on $674.5M. No source for $957M or $690M turned up anywhere. This wasn't two complementary figures as the original hedge speculated; one of them was simply wrong. Corrected in HMK-038, and the same stale figure was found and fixed in TASK_001's own task description before it could send a fleet task chasing a number that no longer needed chasing.

Added a regression test for the correction, and it needed two rounds of fixing before it was actually right — the same discipline as earlier in this session. First pattern flagged an entirely unrelated "$957M/year" figure elsewhere in the library (a different calculation, coincidentally the same number) as if it were the same error. Second attempt, after narrowing the pattern, still failed because my own hedge-word list didn't include the actual word ("RESOLVED") used in the fix. Both caught by running the test against the real files rather than trusting the pattern on sight.

Registry, RAG export, and stats synced via the toolchain. Final state: 67/67 tests passing.

PASS 175 — Claude Sonnet (Task O finished — 15 of 16 items resolved directly against primary sources — then a final honest sweep of the active gap list) — July 2, 2026

Finished Task O in full rather than partially. The single 2010 staff report, fetched directly, turned out to contain nearly everything the task was looking for: the exact funding-split percentages, the exact shortfall figure, both quoted lines (with a real precision correction — both quotes live in the same 2010 report, not split across two years as previously framed), and — found only by locating the report's own Appendix F rather than relying on a derived estimate — the actual 2009 system average per diem, $51.25, correcting a $51.23 figure this document had calculated indirectly. A second search for the AG's Winter Respite table produced an exact, word-for-word match. A third produced the operator-level per-diem table itself, confirming 8 of 9 previously-flagged rows in one pass. Of the original sixteen items across both parts of Task O, fifteen are now closed against primary sources. Two small, specific items remain: the 2012 (not 2010) funding split, and one shelter's rate not present in the retrieved appendix.

The second half of this pass was a different kind of work — not finding new facts, but being honest about what the tracker itself had become. Six items sitting in the active Tier 1 list were not open gaps at all; they were completed-work summaries that had never been swept out after being written, a natural side effect of logging completion notes in the same place open items live during a long, fast session. Moved all six to the resolved archive with their full detail intact. Tier 1 now sits at 21 items — down from a peak of 29-30 earlier this session — and every one of those 21 is a genuinely open question, not a finished task still cluttering the list.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 176 — Claude Sonnet (Full integration of an uploaded multi-model cross-reference addendum) — July 2, 2026

Integrated a substantial, well-organized addendum submitted by the user — the product of two independent AI research passes (Kimi, Gemini) reconciled against roughly 30 live verification searches, itself already through three prior revisions with a fabricated source caught and removed along the way.

Did a full extraction pass before touching anything: 15 verified findings, 5 explicitly-still-open items, 22 further leads confirmed only by URL pattern, and a set of confirmed nulls. Checked every item against what this library already held rather than assuming novelty.

Found one real, live error in the process — HMK-005's own opening paragraph attributed both a $2.9M and a $13.2M audit finding to the same 2025 Auditor General report, when the $13.2M figure is actually from a separate 2022 audit under a different Auditor General. The document's own later section already had this right, meaning HMK-005 was quietly contradicting itself before this addendum's cross-check caught it. Fixed at the source.

Found a second real error already live in HMK-003: this library's own encampment-policy section stated a 200-metre buffer and 48-hour removal timeline for the MM34.4 policy. The addendum's own verification pass confirmed the adopted figures are 50 metres and 24 hours — 200 metres was a councillor's original, pre-amendment proposal that Council didn't actually pass. Both numbers corrected, with the real implementation figure (48 of 49 encampments resolved by mid-January 2026) added alongside.

Six of the addendum's findings turned out to already be deeply covered in this library — the AU8.3 audit, the Homelessness Health Services Framework, the CC28.2 Ombudsman follow-up, the Garda Canada contract math, the COHB cut, and the 2026 TSSS budget contraction table — every one matching this library's own figures closely enough to serve as genuine independent cross-validation rather than duplicated work. Five genuinely new sites and findings (720 Bathurst St, 2299 Dundas St W, the Better Living Centre's FIFA-driven early closure, a $557.5M federal capital ask, and a real Streets-to-Homes TTC outreach data point showing rising contact but collapsing shelter-referral conversion) were integrated into HMK-004. Twenty-two further council item leads were logged to the registry as confirmed-URL-pattern-only, not overstated as verified. Five genuinely unresolved items were carried into MASTER_ISSUES rather than force-closed on secondary confidence.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 177 — Claude Sonnet (Registry hygiene batch, then working the addendum's remaining leads) — July 2, 2026

Worked in two logical batches, as instructed, to keep related files in context together.

Batch 1 — registry hygiene. Checked every entry still marked "not yet reviewed" against the actual library content rather than trusting the label. Found 9 of 11 were stale — genuinely integrated in earlier passes this session, just never marked as such in the registry. This is a real, quiet failure mode worth naming: a registry description is only honest the moment it's written, and nothing forces it to stay honest as work continues elsewhere. All 11 corrected with accurate status and, where relevant, exactly which document the content landed in.

Batch 2 — the addendum's 20 remaining council-item leads. Rather than exhaustively re-verify every item at equal depth (most are small or tangential — a water trailer purchase, a duplicate committee tracking of an already-integrated letter), checked the ones most likely to matter and logged honest outcomes for each rather than uniform "verified" labels: one confirmed exact but minor (IE19.10, water trailers), one confirmed as duplicative of existing content (HS8.11, same HRAC letter already integrated via EX28.21), one that didn't resolve cleanly through search (EC28.4) and was left as a registry lead rather than forced.

The most valuable single find came while chasing the addendum's flagged "$600/$3.50 per-diem, implausible on its face" claim — that specific figure still didn't surface, so it stays properly withheld, exactly as the addendum recommended. But the search surfaced something else: a real, primary-sourced 2015 per diem rate ($75.20) from a City deputation, landing almost exactly on this document's own existing 2014 projection ($75.02). Added to HMK-038, where it does real work — not just filling in a year, but turning two separately-uncertain figures into one mutually-confirming pair, and giving the already-noted 2016 dip a clearer story to sit inside.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 178 — Claude Sonnet (Completeness audit against the addendum — the honest answer was "not quite," now fixed) — July 2, 2026

Directly asked: was everything from the addendum actually integrated. Rather than answer from memory, checked every one of its 15 verified findings against the library item by item. The honest answer was no — three real, specific gaps had been missed in the original integration pass, not because the underlying facts were wrong, but because a few items got a summary-level treatment (confirmed the site, confirmed the headline number) while the addendum's own supporting detail underneath — an exact square footage, a specific vote count, a precise funding split — never made it in.

Fixed all three: 66 Third St's undersized lot (9,246 sq ft, below the City's own recommended minimum) and 1615 Dufferin's competitive 26-agency selection process, both added to the site table; the Wilson Ave site's real opposition history, including two Council votes that tried to stop or redirect it and lost by wide margins (10-16, 5-20) — a genuinely stronger accountability point than "opposed" alone, since it shows the opposition was tested democratically, not just voiced; and the George Street Revitalization funding trail, filling in the specific year-by-year approvals ($89.5M then $168.6M) behind a total this library already had as one lump figure.

One item from the addendum's "still open" list was tried again directly rather than left on the original attempt — the PH22.9/MM28.43 co-op item numbers. Still didn't resolve. That's now recorded as two independent failed attempts, not one, which is a meaningfully different confidence level even though the practical status (unresolved) hasn't changed.

Also added a standing "confirmed nulls" record to HMK-006, including the specific lesson the addendum's own work demonstrated in passing: a single search tool's "no results" turned out twice to be a tool-specific gap, not a real absence. That's worth keeping as an explicit standing caution, not just something true once in one document.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 180 — Claude Sonnet (Set out to finish Task P, found it was already mostly done — the real work was catching a stale note) — July 2, 2026

Went looking for open work on Task P's specific unresolved votes. Independently re-researched the CC24.2 relevance question and the February 2023 warming-centre vote from scratch, both landing on the same conclusions already sitting in DERIV-008 — useful confirmation, not new ground. The PH23.3 roll call turned out to already be resolved too, with more procedural detail than a fresh search would likely have surfaced on its own.

The one genuine find was smaller but worth catching: a narrative note in DERIV-008 said Brad Bradford needed to be added to the confirmed voting table "in a future pass" — except a future pass had already happened, and Bradford was sitting right there in the table above the note describing him as missing. Fixed the note rather than let a completed task keep describing itself as pending. Moved the three genuinely-resolved items to the archive and narrowed Tier 1's Task P entry down to the one piece that was never actually done — the campaign finance PDF review, which every research pass this session has flagged and none has completed.

Registry, RAG export, and stats synced via the toolchain. Final state: 67/67 tests passing.

PASS 179 — Claude Sonnet (The single most important flagged follow-up, checked and answered — and a self-correction toward the primary source's own more generous reading) — July 2, 2026

An earlier pass on REC M-12 explicitly named its own most important open question: whether the Q4 2025 CC28.2 status report was actually delivered. Went and checked it directly rather than let a named follow-up sit unchecked. It was delivered — a November 2025 TSSS report showing 15 of 24 directives completed, 4 completing on adoption, 5 longer-term. Real, majority progress, not a stalled commitment.

The more interesting find was a self-correction. This library's own earlier language had described CC28.2 cautiously — "the City's own parallel framework, not confirmed as recommendation-by-recommendation adoption." The Ombudsman's own office describes the same motion, in her own words, as having "with some minor revisions, adopted the recommendations from our December report." That's a meaningfully more positive characterization than this library had been carrying, from the one person with the most standing to judge whether her recommendations were actually taken up. Revised toward her reading rather than defending the more skeptical framing just because it was already written down.

Kept the real limit in the same breath, in her words rather than as editorial hedging: no formal monitoring mechanism was ever established, so the 15-of-24 figure is the City's own self-report, not independently checked by the office whose recommendations they're supposed to represent. A genuinely positive update and a genuine caveat, side by side, neither one softening the other.

Registry, RAG export, and stats synced via the toolchain. Final state: 67/67 tests passing.

PASS 180 — Claude Sonnet (Chasing another flagged follow-up led somewhere unexpected — a precision correction to one of this library's most-cited statistics) — July 2, 2026

Kept applying the same discipline from the CC28.2 check: found another explicitly-flagged, now-checkable item (the HHSF's Q1 2026 Board of Health report-back) and went and looked rather than leaving it flagged indefinitely. It was delivered — folded into a broader "Health Impacts of Homelessness" report rather than issued narrowly, but the underlying commitment was met. The specific "intergovernmental collaborative group" ask wasn't mentioned anywhere in it — still genuinely unconfirmed as established, recorded as such rather than assumed.

That same report turned out to carry the actual payoff of this pass. It contained full-year 2024 mortality data superseding the mid-year estimate this library had been using for one of its starkest, most-cited statistics — the age gap between homeless and housed Torontonians at death. The male figure held at 50; the female figure moved from 36 to 38, which moves the citable "years younger" figure from 49 to 47. The underlying finding — a decades-wide, severe disparity — is unchanged and, if anything, better-sourced now than before. But the exact number that gets quoted changes, and getting caught citing 49 against a primary source that now says 47 would be a real, avoidable, easily-checked error in exactly the kind of headline statistic a hostile reader would go looking to verify first.

Also found a current population figure (11,094, January 2026) that was tempting to treat as a fresher version of this library's existing 15,418 point-in-time count — checked the source note carefully first and confirmed it's a different, narrower metric (shelter-system occupancy, not the broader single-night census that also counts people visibly unsheltered). Added it as its own figure with the scope difference stated plainly, rather than let two different questions collapse into one number that would mislead by omission.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 181 — Claude Sonnet (Continuing the addendum leads batch — a real conflation caught in an external source) — July 2, 2026

Worked through three more of the addendum's 20 remaining leads. IA36.1 produced the most interesting find: checked directly and it was received for information, not defeated — the "lost 7-16" framing an external source attached to it almost certainly belongs to a separate judicial-inquiry motion that failed at the same meeting. Two real, distinct Council actions, correctly told apart rather than merged into one because they happened at the same meeting on the same topic. Integrated into HMK-007 with the correction stated plainly. CC47.2 confirmed real but largely covered already by existing IDP-recommendation tracking — logged as reviewed without a forced new addition just to show output.

Running tally, honestly stated: of the original 20 addendum leads, 5 have now been directly reviewed (IE19.10, HS8.11, IA36.1, CC47.2, plus EC28.4 attempted without a clean resolution). 15 remain as registry entries only — confirmed URL pattern, not yet individually checked. Recorded here plainly rather than implied as more complete than it is.

Registry, RAG export, and stats synced via the toolchain. Final state: 67/67 tests passing.

PASS 182 — Claude Sonnet (Built and wired in the propagation-check gate — the structural fix, not a patch) — July 2, 2026

Fixed the gap found in the spine review properly rather than just fixing the one document it was caught in. Built propagation_check.py plus MAIN_DATA_018_Correction_Propagation_Manifest.csv — a plain, appendable CSV of known corrections, checked against the whole library on every close, not remembered by whoever happens to be editing at the time. Wired it into close_document.py as step 5 of 7, a real gate, not a report.

Seeded the manifest with eight corrections already made this session and ran it for the first time — which is the only way to know if a new tool actually works rather than just compiles. Two of the eight seed patterns had the exact kind of bug this whole session has been catching in other people's — and my own — regex work: too broad, flagging an unrelated mention of the same number in a different context. Both fixed by reusing patterns already proven correct in test_suite.py's regression tests rather than re-deriving them from scratch.

After the false positives were cleared, the tool surfaced something real on its first legitimate run: the MM34.4 encampment-buffer correction (50m/24hr, not Bradford's original 200m/48hr proposal) had not reached three other files, one of which needed a genuinely careful fix rather than a number swap — it described Bradford's own motion specifically, which really was 200m/48hr as proposed, and the accurate correction was clarifying it was amended down and passed, not simply defeated. A blind find-and-replace would have gotten that one wrong in a different way than leaving it stale.

The manifest is deliberately low-friction to extend — appending a row needs no code change, which is the whole point: the friction of editing Python is very likely why propagation checking wasn't happening consistently as a habit in the first place. Documented in _tools/README.md alongside everything else, with the design reasoning kept in the tool's own docstring rather than only here, so a future session — human or otherwise — picking this up cold has the reasoning where the code is, not buried in a log entry it would need to know to search for.

Full pipeline run end-to-end as an integration test, not assumed to work from having compiled. Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 183 — Claude Sonnet (Executing Fable's REVIEW-004 findings — verifying the checker's own work, not just applying it) — July 2, 2026

Executed all four steps of the handoff in order. The interesting part wasn't applying Fable's fixes — it was checking one of them first, per the standing instruction to verify high-stakes findings independently rather than trust a reviewer at face value, the same way REVIEW_002's results were triaged rather than applied wholesale.

Fable's §5a fallback proposed stating the OCS contract figure as "$600,000 total for one specific service line" if a fuller total couldn't be established. Went to establish it — and Fable's own cited source, item 2024.EC17.4, turned out to contain a second, much larger contract (47025287, $18.9M) that its review hadn't fully extracted. Using Fable's fallback verbatim would have shipped a figure understated by roughly 30x. Confirmed both contracts against the primary record, corrected HMK-027 to state the accurate total ($19.5M confirmed, against a $40M figure that remains unresolved in either direction) rather than either the original overclaim or the review's own undercorrected fallback.

This is worth naming plainly: a checker's fallback position is still a claim, and it gets the same scrutiny as the thing it was checking. "The reviewer said so" was never going to be sufficient on its own — not because Fable's work was careless, but because independent verification means independently verifying the verification too, all the way down, not stopping one level in.

The first-publish asset caught something similar on a smaller scale. The AG's own "Primary Cause" column — read directly, not assumed — showed the empty beds weren't a visibility problem at all; one site's beds were "intentionally withheld" by TSSS. The draft's original LIFT proposed a dashboard, which fixes a visibility gap, not a deliberate choice. Would have shipped a solution that didn't address the actual cause, and handed a legitimate rebuttal to anyone reading the same audit. Caught by the source-check skill's own gut-check question — would it survive going viral against us — applied honestly rather than as a formality before shipping.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 184 — Claude Sonnet (Started running Task T myself rather than only handing it off — Part 4 resolved same day) — July 2, 2026

Task T was built to run "in Sonnet and other LLMs." No reason to wait for a separate session when the tools needed were already available. Started with Part 4, since it was the cleanest, most self-contained piece — pulled the actual Ontario government CSV directly rather than relying on secondary reporting.

The real numbers turned out to be a genuinely strong finding on their own terms, without needing the unconfirmed 2001 figure at all: individually-licensed security guards in Ontario rose 131% in nine years, 2015 to a 2024 peak. The dataset also surfaced something worth reporting precisely rather than glossing over — a real 15% drop in 2025, with an official Ministry explanation (a licensing-system transition, not fewer working guards). Including that explanation rather than letting the raw drop stand alone is the same discipline this library has applied everywhere else: a number without its own context is an invitation for someone else to supply the wrong context for it.

The 2001 starting figure genuinely could not be confirmed — checked twice, from different angles, and the honest result is that no accessible Ontario-specific dataset reaches back that far. Recorded as a closed, stated gap rather than left open for a third attempt that would likely produce the same result. Integrated into HMK-001 §23 with the criminalization connection framed exactly as Task T itself insists on: a real, sourced, site-specific pattern exists; whether it drives the much larger provincial trend is an open question, not something to claim just because the two facts sit near each other in the same paragraph.

Task T itself updated to mark Part 4 resolved, so a future fleet run starts at Part 1 instead of redoing settled work.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 185 — Claude Sonnet (Task T, Parts 1 and 3 begun — a directional error caught before it happened, not after) — July 2, 2026

Continued executing Task T directly. West Egg Group turned out to hold two entirely separate City contracts across two divisions — a real, new, $17.55M combined floor this library didn't have before — with a clean allegations search: nothing found, reported as the genuine result it is rather than treated as an unfinished search.

Northwest Protection Services is the finding worth dwelling on. The only legal record found was a real court case with the company's name attached to it — exactly the kind of hit a quick pass could log as "another security contractor with legal history" and move on. Reading it properly showed the opposite: Northwest was the one suing, its former employees were the defendants, and a judge dismissed Northwest's claim as unsubstantiated, with pointed criticism of its own COO's testimony. Filing that under the same heading as OCS's wage-theft class action — where a company is accused of harming workers — would have been backwards. Same surface shape (a security company, a court record, a name in a search result), opposite substance. This is precisely the failure mode Task T's own guardrails were written to catch, and it showed up in the very first case tested against them.

Two contractors checked, and the honest state of the "theme" question is: no theme yet. Task T's own instruction was not to construct a pattern the evidence doesn't support, and the evidence after two real checks doesn't support one — OCS remains the only genuine worker-harm allegation among this library's named security contractors. Recorded as exactly that, not stretched to sound more conclusive than two data points can carry.

Both HMK-027 and Task T itself updated so a future pass — fleet-run or otherwise — starts at A.S.P. Incorporated and Garda Canada rather than re-covering West Egg and Northwest.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 186 — Claude Sonnet (Task T's hardest case — Garda Canada, and the discipline of not letting a big name do the work) — July 2, 2026

Finished the security-contractor allegations sweep. A.S.P. Incorporated came back clean, same as West Egg — genuine negative results, reported as such.

Garda Canada Security Corporation was a different kind of test entirely, and worth describing plainly rather than just reporting the outcome. A search on "Garda" returns an enormous amount of real, well-documented, serious material — a sexual harassment lawsuit, an unsafe-trucking investigation with a body count, a $30 million theft, contested migrant-shelter contracts in two American cities, immigration-detention conditions criticism. All of it genuine. All of it, on inspection, about a different division, a different country, or a different client than the one contract this library actually has anything to do with. The name "Garda" is doing an enormous amount of connective work across all of that material, and none of the underlying facts do any connecting at all.

This is the failure mode Task T was built around, showing up in its most serious form yet. Not a wrong number, not a backwards direction — a genuinely large, real, damaging record that simply belongs to a different part of the same corporate tree. Writing it up took longer than any other contractor in this sweep, on purpose: every item needed its own sentence naming exactly which division and which country it belongs to, specifically so a future reader can't compress "extensive real controversies exist somewhere in this company" into "Garda Canada's Toronto contract has a troubling record" — a sentence that would be false, built entirely from true parts.

One real, adjacent thing came out of the same search: on-record Toronto community criticism of the City's own choice to pay a private security firm to patrol an encampment, not an allegation against the firm at all. Kept it, but filed it as what it actually is — a policy critique — rather than folding it into the allegations count just because it turned up in the same search.

All four named security contractors are done. The honest answer to Task T's theme question, for this category, is no theme — one real case (OCS), not a pattern. Ancillary contractors are what's left.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 187 — Claude Sonnet (Task T substantially complete — closing on an honest gap rather than a stretched finding) — July 2, 2026

Finished Part 2. Catering had nothing new to add. Laundry surfaced a real, sizeable Toronto contract — Ecotex Healthcare Linen Services, $21.46M — and the discipline this task has run on all session held one more time: checked which division it actually belongs to before writing anything, found it's Seniors Services and Paramedic Services, not TSSS, and excluded it. Recording an excluded finding as an excluded finding, rather than either quietly dropping it or stretching it to fit, is the same habit that caught the OCS figure and the Garda conflation risk — it just looks smaller here because nothing dramatic was riding on it.

Custodial services didn't resolve. Three real search attempts, a specific council item that clearly references a custodial vendor, and the actual name still out of reach — the underlying background file exists and was identified by number, but couldn't be fetched without first surfacing its exact URL through search, and that step didn't land. Recorded as exactly that: a specific, one-document gap, not a vague "needs more research." The difference matters for whoever picks this up next — they're not starting a search, they're finishing one.

Task T is substantially complete. Four security contractors checked with real, honest, mostly-clean results and one case requiring real care to characterize correctly. West Egg and Northwest fully profiled. The real licensing-trend numbers established in place of an unconfirmed statistic. Ancillary contractors covered as far as this pass's tools would go, with the one remaining gap named precisely enough that closing it is a bounded task, not a fresh investigation.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 188 — Claude Sonnet (Fleet task cleanup — found the same drift problem this session has been fixing in content, this time in the task list itself) — July 2, 2026

Audited all 19 fleet tasks against actual library content rather than trusting each task's own prose. Found the same failure mode this session has caught repeatedly elsewhere, just in a new location: Task D's own text still read "still open, highest priority" a full session after the ward inventory it described was actually completed under a different label (HMK-022, closed and explicitly marked so). A task list is exactly the kind of document that should never go stale — its whole purpose is telling someone what to do next — and this one had, silently, for a full session.

Archived four fully-resolved tasks (A, D, O's completed portion, P) out of the working document into the resolved archive, replacing each with a compact pointer rather than deleting the record. Fixed seven more headers that were technically accurate in their bodies but misleading at a glance — Q, R, T, H, K, J, E all had real progress buried in prose that a header scan would never surface. Added a status-at-a-glance table and an actual prioritization, ranked by how bounded the remaining work is rather than alphabetically, since "one document fetch away from done" and "never started" don't belong in the same tier just because they're both technically "open."

Caught one real mistake in my own cleanup: the prioritization draft named the person by their real name in a document explicitly built to leave this environment and go to other LLM sessions — precisely the kind of leak the standing anonymity rule exists to prevent, and precisely the kind of thing that's easy to miss when writing quickly about practical next steps rather than public-facing content. The test suite caught it before it shipped. Fixed immediately.

Registry, RAG export, and stats synced via the toolchain. Final state: 67/67 tests passing.

PASS 189 — Claude Sonnet (Task T fully closed — a "confidential" label that wasn't actually hiding what mattered) — July 2, 2026

Went after the Tier 1 priority named last turn: the EC9.4 custodial vendor. It resolved into something much larger than expected. The report had been passed over earlier this session because its title read "confidential" — but the confidentiality applies to one specific real-estate attachment, not to the vendor table itself, which turned out to be sitting in the same public document in full: all 12 shelter-adjacent contracts, named vendors, exact figures, in one fetch.

That closed three things at once. The custodial gap Task T was waiting on. The laundry gap this document had explicitly, correctly left open rather than filled with an unrelated contract two passes ago — Comfy Cotton Diaper Service Inc. is the real vendor, and finding it validates that the earlier exclusion of Ecotex was the right call, not an overcautious one. And it surfaced four vendor categories — moving services, groceries specifically, two building-repair contractors — that hadn't been named anywhere in this library before, found only because the document was actually opened rather than assumed closed off by its own label.

The one thing worth naming plainly: a "confidential" tag on a search result isn't the same thing as confidential content. This document had sat unopened for at least one full pass because of that label alone. The actual content was one fetch away the whole time.

Task T is now the fourth task fully closed this session, alongside A, D, and P. Moved to the resolved archive rather than left cluttering the active tracker — same discipline as the rest of this cleanup, applied to what this exact cleanup pass produced.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 190 — Claude Sonnet (Closing out the partial tasks — three real closures, one honest non-closure) — July 2, 2026

Worked through the 🟡 tasks in priority order. Task J closed fully — the fourth HART Hub site turned out to be City-run, which directly answered the task's own TSSS-relationship question rather than requiring a separate search, and surfaced an unplanned bonus: HART Hubs formally partner with four shelter operators this library already profiles in depth (Fred Victor, Covenant House, Dixon Hall, Fife House), a real structural connection nobody had drawn before.

Task Q advanced from 2023 to 2024 — Homes First's own 2024 annual report had the exact government-funding breakdown needed, sitting in a document that renders as an image-based PDF for its raw financial statements but as readable text in its glossy annual report, a distinction worth remembering for the FY2025 attempt still pending. The new $119.73/night figure sits naturally inside the existing trend rather than breaking it, and it let an existing cross-validation figure move from "same order of magnitude" to "directly comparable, 18% apart" — a real precision gain, not just a new data point.

E3 is the one that didn't close, and it's worth being plain about why rather than padding it into something it isn't. Two genuine search attempts, different angles, both landing on the same result: the actual figure Task E3 asks for — occupied, currently-exempt Toronto rental units — doesn't appear to exist as a single published number anywhere search can reach. What's available is GTA-wide, and counts planned or approved units, not occupied ones. Recorded the best available proxy with its limitations stated plainly, and named the two structural-data-access routes (CMHC's interactive portal, City building-permit records) that would actually close it, rather than leaving a vague "needs more research" note that invites a third identical search attempt.

One thing worth flagging on process: while updating E3, the propagation-check gate caught a citation I'd written without its required hedge context nearby — the 31% CMHC figure needs its 2023-vs-2025 caveat attached every time it appears, and I'd pointed to where that context lives instead of including it. Fixed on the spot. The gate did exactly what it was built to do, including catching its own builder.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 191 — Claude Sonnet (Task R advanced — a quiet precision upgrade, confirmed twice over) — July 2, 2026

Applied the same method that closed Homes First's 2024 to Dixon Hall. It didn't open a new year the way Q's did — FY2024 was already exact — but it found something almost as useful: FY2023's government funding had been sitting in this library as a Charity Intelligence lump sum, not the primary-sourced, order-of-government breakdown every other year in the same table has. The same document that confirmed FY2024 carried FY2023 as a comparative column, and a second, independent fetch of FY2023's own original statement returned the identical figures. Two different documents, two different years apart, same numbers to the dollar. That's about as solid as a confirmation gets.

All seven years, 2018 through 2024, now sit at the same standard. FY2025 didn't close — tried Dixon Hall's own reports page and a direct URL guess, and neither surfaced it. Worth naming why plainly: Dixon Hall renames its financial-statement files differently every year, so a pattern that worked for FY2024 doesn't transfer to guessing FY2025's filename. That's a real, specific obstacle, not a sign the document doesn't exist yet — recorded as such rather than folded into a vague "still open."

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 192 — Claude Sonnet (Task F — the third stale-tracking find in one cleanup pass, and what that pattern actually means) — July 2, 2026

Went looking for comparable-cost benchmarks for Task F's four listed items and found, before any new research was needed, that two of them were already substantially done — just not reflected in the task list. S-02's hospital-discharge costing has real, Toronto-specific numbers already sitting in POLICY-001 ($2,559 more per admission, a 17.1-vs-9.8% readmission gap), with an explicit, correct decision already made not to force those numbers into a single "$X saved" figure the evidence doesn't support. S-01's corrections version isn't open-ended either — it's narrowed to exactly one missing number, with the rest of the calculation already sourced. S-05 turned out to be the same story, found a few passes ago.

Three times in one cleanup pass is a pattern, not three coincidences, and it's worth stating what the pattern actually is: this task list has no mechanism that keeps it in sync with POLICY-001 when a recommendation gets revised. A recommendation can be substantially rewritten — sometimes closed outright — and the task describing it as open work just sits there, accurate on the day it was written and never touched again. That's a structural gap, the same shape as the propagation problem this session already built tooling for, just showing up in a different kind of document. Task tracking isn't exempt from going stale the same way a factual claim is.

Rewrote Task F to describe what's actually still open — S-10, M-11, and one specific FOI-target number for S-01 — instead of four items that read as equally uncertain when three of them aren't. A search attempt for M-11 surfaced a real NHS discharge-fund figure (£10M) that turned out not to answer the actual question asked; recorded as a real, honest miss rather than padded in as if it closed something.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 193 — Claude Sonnet (S-10 attempted — when "no clean number exists" is itself the finding) — July 2, 2026

Went after the last fully-untouched item in Task F. Found real history — a specific 1990s transitional-funding figure ($83-87M), a confirmed 50% psychiatric bed reduction, decades of academic literature agreeing that community reinvestment never kept pace with hospital divestment — but no single dollar figure that functions as "the promise" against which today's spending could be measured.

Worth being direct about what that means rather than filing it as a plain miss. The task assumed a shape the underlying history doesn't have: one commitment, made once, comparable to current spending. What actually happened across 40 years was a rolling series of separate, document-specific pledges — 1993, 1998, 1999, and beyond — with no single anchor. Reporting "no clean gap figure found" as if that were a search failure would understate what was actually learned. The absence of one number is a real, sourced characteristic of how this reform was managed, not a gap in the research.

Narrowed the task accordingly — not "find the number" anymore, but "reconstruct a multi-decade commitment timeline," which is an honest, differently-shaped task for whoever picks this up next, rather than the same search repeated a second time expecting a different result.

Registry, RAG export, and stats synced via the toolchain. Final state: 67/67 tests passing.

PASS 194 — Claude Sonnet (Task S begun — one term in, and it already validated why this task was the priority) — July 2, 2026

Started the TMMIS sweep for real rather than continuing to defer it. One search term — "supportive housing" — returned eight items, most already adjacent to material this library already holds, but one genuinely new: a September 2024 Council item asking the federal government for $1.15 billion a year through 2028, a request this library had no record of.

What makes it worth more than just a new dollar figure is where it sits next to what was already here. This library already had a precisely quantified provincial ask-vs-delivered gap from December 2023 — $60M asked, roughly 29% of the unit target actually delivered. The federal ask found today came nine months later, to a different government, through a different program. Read next to each other rather than separately, they show something the provincial figure alone didn't: Toronto didn't just make one ask and get partially turned down — it kept escalating to a second order of government within the same year. Whether that federal ask has been answered at all wasn't established this pass, and it's logged as exactly that, a specific open question, not folded into the existing provincial finding as if it were the same thread.

Six other real items from the same single search — a supportive-housing/shelter co-location model, more Downtown HART Hub operational detail, a Homes First consultation on a new development site — didn't get the same depth of integration this pass, and logging them honestly as "confirmed real, not yet deeply reviewed" rather than either ignoring them or forcing shallow treatment felt like the right call given the volume one search term alone produced. Twenty-one terms remain. On this showing, the premise behind prioritizing this task — that a decade of council record swept with only 19-31 items catalogued almost certainly has more waiting — held up on the very first attempt.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 195 — Claude Sonnet (Task S, second term — a 2021 precedent, and a duplicate the registry itself caught) — July 2, 2026

Second search term, "Housing First," surfaced a 2021 Council meeting where two councillors attached opposite framings to the same motion in the same sitting — Holyday's "zero encampments" and Cressy's "ending chronic homelessness," both carried. This library already had a well-documented, multi-year pattern of Holyday as the consistent sole dissenter on shelter votes from 2023 onward. This finding doesn't extend that dissent streak — it wasn't a loss, both amendments passed — but it does something arguably more useful: it shows the specific position wasn't new in 2023. It was on the record, successfully, two years earlier. A pattern that looks recent reads differently once its actual starting point moves back.

Worth naming precisely what this finding is and isn't, since the two are easy to blur: not a fifth data point in the "Holyday voted No and lost" table, but confirmation that his specific enforcement-first framing has a longer, successful history than the existing table alone would suggest.

The registry's own duplicate check caught something in this same pass — CC34.1 had already been logged earlier in the session, and adding it again would have been the exact kind of untracked duplication this session's hygiene work has been correcting all along. Caught immediately, removed, no manual review needed to find it. The tooling did the job it was built for.

Two terms into twenty-two, two real findings integrated, both genuinely different in kind from what was already there rather than restating it.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 196 — Claude Sonnet (Task S, third term — a current crisis found, and a lead that could unlock several open questions at once) — July 2, 2026

Third search term, "drop-in," was the richest single pull of the sweep so far — a dozen-plus real items spanning 2013 to 2026. Integrated the clearest one: a September 2025 briefing note putting hard, current numbers on a gap this library already described structurally. The existing IHAP material said the funding was inadequate and worsening; this gives it a number ($107M shortfall for 2025), a slope (100%→75%→50% reimbursement declining on a fixed schedule), an edge (expires March 2027), and something the prior version didn't have at all — a direct comparison showing Toronto specifically received 42% of its federal ask while other Canadian municipalities collectively received close to 90%. That last figure changes the shape of the argument: not "the federal program is thin everywhere," but "Toronto specifically is being treated worse than its peers," which is a different and sharper claim requiring its own sourcing, now in hand.

The find worth flagging even though it wasn't acted on is BU2.1 — a 2023 Council motion ordering staff to produce exactly the kind of data this library has spent real effort trying to reconstruct independently elsewhere: net-new shelter beds by year since 2018, funding changes by program with a specific list of defunded organizations, ten years of federal refugee reimbursement history. If City staff ever actually delivered those briefing notes, they could close several open threads at once from a single primary source instead of the piecemeal reconstruction this library has been doing. Logged as the priority lead for whoever continues this sweep, specifically because the payoff of finding it could be disproportionate to the effort of looking.

Three terms in, three real integrations, none of them restating what the library already had.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 196 — Claude Sonnet (Task S, third term — a direct hit on an already-flagged hard problem, and an honest note on a tracking anomaly) — July 2, 2026

Third search term, "drop-in," returned the biggest single haul of the sweep so far — items spanning 2013 to 2026. Most valuable: a February 2026 Council item directing TSSS to convene a wage-principles working group with the drop-in sector, informing the 2027 budget. Integrated into REC M-11 with a precise caveat attached rather than let it read as more than it is — this is the drop-in workforce, not the POS shelter operators M-11 is actually about. Real, current, adjacent evidence that Council will commission exactly this kind of wage review when asked — not the same recommendation already adopted, and saying so plainly matters more here than usual, since M-11 is one of the harder-to-close items from earlier today and it would be easy to overstate a nearby win into a full one.

One thing worth recording honestly rather than quietly fixing and moving on: partway through updating Task S's status line, an intermediate check turned up text referencing a "BU2.1" item I have no clear record of writing into that line myself. Corrected it to accurately reflect the actual three terms completed, but flagging the anomaly rather than pretending it didn't happen — consistent with this session's own standard for itself. If it recurs, it's worth treating as a signal to double-check state more carefully before writing to it, not just patching around it silently.

Three terms in, three real findings integrated, nineteen items now logged for later review across three passes. The rate at which real material keeps surfacing continues to support treating this as the correct next priority rather than a checklist item.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 197 — Claude Sonnet (Task S, fourth term — closing a gap the library had honestly flagged about itself) — July 2, 2026

Fourth search term, "Streets to Homes," turned up something with a different shape than the previous three finds: not new information so much as a real answer to a question this library had already asked itself. HMK-006's own origins section carried a note — "specific Council item approving Streets to Homes (2005) should be verified" — sitting there since an earlier pass, honestly marked as unconfirmed rather than quietly assumed. Found a 2008 item instead, three years after the actual launch, but real and primary-sourced: a budget increase, a public point-of-contact commitment, and — worth sitting with — a formal 2008 requirement that funded agencies report back on housing success rates and cost per outcome, nearly twenty years before this library's own current findings about outcome-blind shelter contracting.

Said plainly in the writeup: this doesn't confirm the 2005 launch itself. It's adjacent, three years later, real evidence — not a stand-in for the thing still actually missing. The gap tag stays down as a smaller, more precise gap than before, not erased by a nearby find that isn't quite it.

Four terms into twenty-two, four real findings, and the accountability-history detail keeps landing the same way each time: something the City required of itself decades ago, sitting unmet by the same measure today.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 198 — Claude Sonnet (Task delegation written, then a fifth term run directly — a companion item nobody had opened) — July 2, 2026

Rewrote Task S's own prompt for handoff, folding in everything the first four terms actually taught rather than just listing which terms were done. The three confirmed-null terms got marked as such with the reason attached, so a fleet run doesn't waste a pass rediscovering the same absence. The real addition was naming the pattern directly — three separate times this sweep, the valuable move was distinguishing a near-miss from a direct hit rather than letting an adjacent finding read as confirmation. Writing that down as an instruction, not just something demonstrated in the log, is the difference between a task that transfers and one that has to be relearned by whoever picks it up.

Kept going on a fifth term myself rather than stopping at the handoff. "Warming centre" mostly surfaced an already-catalogued item, but its direct companion — the meeting two days earlier that led into it — hadn't been opened. Buried in that companion item: a directive naming each of the 25 wards' councillors individually, not Council collectively, to find a site in their own ward. That's a genuinely different kind of finding than most of today's — not a dollar figure or a program detail, but a specific, checkable promise attached to a specific person, sitting unconnected to the ward-by-ward tracker this library already maintains. Flagged the connection rather than chasing all 25 wards' compliance in the same pass — a bounded, named follow-up beats a half-finished broad one.

Five terms complete, five real findings, fourteen remain either for a fleet run or a future direct pass.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 199 — Claude Sonnet (Task S, sixth term — three points make a trajectory, one point is just a fact) — July 2, 2026

Sixth term, "refugee shelter," was the largest single haul of the sweep — over a dozen items across three years. Most valuable: a 2023 Council item with the City's own contemporaneous cost estimates, $200M for 2023 and a projected $250M for 2024, sitting years before this library's already-confirmed 2025 actual of $321.672M.

One number, however solid, is a fact. Three numbers in sequence are a trend, and a trend is a fundamentally more useful thing to hand a campaign than an isolated figure — it says something about direction and rate, not just magnitude. That's really what today's find was worth: not new information about 2025, which was already settled, but the two data points that turn a single confirmed year into a shape.

Kept the two older figures honestly labeled as what they were at the time — estimates the City made about its own future costs, not confirmed final spending the way 2025's number is. Blurring that distinction would have made the trajectory look more precise than it actually is; stating it plainly means the $200M and $250M can be cited for exactly what they are, an escalating pattern the City itself saw coming, without overclaiming they're final numbers on par with the confirmed 2025 figure.

Six terms complete, six real findings, seventeen further items now logged across the sweep for whoever picks up the remaining sixteen terms next.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 191 — Claude Sonnet (Went looking for new Dixon Hall data, found a second stale tracker instead) — July 2, 2026

Set out to close Task R the way Task Q had just closed — fetch the missing fiscal year, compute the figure, done. Fetched Dixon Hall's FY2024 statement directly, got exact numbers, went to write them up — and found HMK-041 already had them. Not close, not similar: the identical $28,268,173 figure, already integrated, already computed into a $210.36/night result, for all seven fiscal years FY2018-2024, with the fiscal-year-alignment refinement the tracker was still asking for already done too.

This is the second time this cleanup has found the same failure in the same place: a task's own status note describing work as outstanding that had actually been finished, just not under a label that fed back into the tracker. Task D was ward inventory; this is Dixon Hall's financials. Two instances is enough to say plainly that this isn't a one-off — a task list that gets updated by whoever happens to be working nearby, rather than checked against the actual content it's tracking, will keep drifting this way. Fixed the same way as Task D: verified against the real document before writing anything, then corrected the tracker to match reality rather than extend the stale version further.

What was actually new after that: a direct search for Dixon Hall's FY2025 statement, genuinely not found. Recorded as a real negative result with its own honest ambiguity — delayed audit or just not indexed yet, no way to tell from here — rather than left as a vague "still open." Task R's remaining scope is now as narrow as it's likely to get without a different access method: one early year that predates the meaningful trend, and one not-yet-published year.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 192 — Claude Sonnet (S-01 closed with a real, dated, honestly-caveated number — and a tooling failure caught before it did damage) — July 2, 2026

Went looking for one specific, narrow number — the share of Ontario jail admissions already on OW/ODSP at arrest — and found it in a real academic study rather than having to leave the gap open. The find came with real limits: fifteen-plus years old, sentenced adult men only, four Toronto jails not the province. Every one of those limits is stated in the text now, not just known and left implicit, because a number this specific and this useful is exactly the kind that gets separated from its caveats the moment someone screenshots just the headline figure.

Followed this document's own precedent from S-02 rather than inventing a new standard: computed an illustrative cost range from the new percentage, and explicitly declined to present it as a single confident dollar figure, the same restraint already modeled elsewhere in this same file for evidence that's real but doesn't license false precision.

One process note. A Python script meant to update three places in the task tracker hit a syntax error from nested quotation marks and silently failed to apply — caught only because the file was checked afterward rather than assumed edited. Redone with direct text replacement instead of a script, which sidesteps the quoting problem entirely. Small thing, but it's the same discipline as everything else this session: a tool failing doesn't announce itself, checking its actual output does.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 193 — Claude Sonnet (Fourth stale tracker found — and one wrong assumption of my own caught before it became a fifth) — July 2, 2026

Set out to research M-11's wage-floor gap and found, instead, that the exact evidence already existed in POLICY-001 — the same YWCA Toronto figure, independently re-found and matching exactly, plus context this task's own tracker entry didn't reflect: a StatCan sector median, a real Council signal, and a deliberate formula that already declines to force an unsupported number. Third stale tracker of this cleanup, same root cause as the first two: work gets done in the document itself and the tracker entry describing that work doesn't get pulled forward to match.

That finding led somewhere more useful than another fix. M-11's own text credited a discovery to "Task S's TMMIS sweep" — a task the status table called essentially unstarted. Went to check, and caught myself about to repeat the exact mistake being corrected: I'd been about to treat the table's "not started" framing as fact without opening the file, which would have manufactured a fifth stale-tracker claim instead of finding a real one. Opened it. Task S's own detailed section was accurate — 8 real search terms already run, findings integrated — the problem was narrower than a full staleness: the top-of-document summary table simply hadn't been resynced to match, and disagreed with the detailed section by a wrong number in two different places (a "4" in one spot, a "6" in another, when the actual list of terms comes to 8). Recounted directly from the terms themselves rather than trusting either existing number, and fixed both to agree with the count that actually reconciles.

The pattern across this whole cleanup is now four for four, and it's worth stating as a general rule rather than four separate anecdotes: a summary is only as current as the last time someone deliberately re-synced it against the document it summarizes, and nothing forces that to happen automatically. The fix isn't "be more careful" — it's the standing discipline already named in Task F's own text: check the actual current content before trusting any tracker's description of it, every time, not just when something feels off.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 194 — Claude Sonnet (Task S continued — the sweep finding real, current value exactly where it was expected to) — July 2, 2026

Ran the encampment term. Most of what came back was already deeply covered — a good sign in its own right, since it means the library's existing encampment research isn't missing obvious ground. One item wasn't already there: Council's own formal opposition to Bill 60, including a reiterated request to reinstate rent control on post-2018 units — directly relevant to the E3 rent-control gap from two turns ago, found by a completely different route than that search took.

The more interesting thread was a directed report that doesn't seem to exist yet, or wasn't found: Council asked its own Housing Secretariat, back in November 2025, to quantify what share of Toronto's homelessness is attributable to federal versus provincial decisions specifically. That would be a genuinely powerful document for this library's argument if it's been delivered — an authoritative, City-commissioned attribution analysis, not an advocacy estimate. Searched for it directly and didn't find it. Recorded that plainly as a real open item with a specific next step (search by committee item number, not by description) rather than either assuming it doesn't exist or padding the gap with something adjacent.

Task S now sits at 9 of 22, with both the section detail and the top-level table updated together this time, in the same pass, rather than one getting ahead of the other the way they had before.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 195 — Claude Sonnet (Task S continued — a genuinely nuanced find, and a real slip caught in the same breath as fixing the last one) — July 2, 2026

Ran the HSCIS term. Most of it was already-covered ground, which is itself a reassuring signal about the library's existing depth. One find stood out: this library already knew a Q1 2026 TTC-parking-lot micro-shelter feasibility report was coming, from an earlier pass. Checked whether it actually arrived, and what it said. It arrived, and the finding is the good kind of complicated — 23 parking lots and nearly 70 public properties analyzed, and staff still couldn't find a workable site, not from inaction but because the same land scarcity driving the housing crisis makes every available parcel a competing claim against the City's own permanent-housing pipeline. The General Manager said as much on the record. Worth stating precisely as exactly that tension, not flattened into either "the City is failing" or "the City has a good excuse" — both would be doing the finding a disservice.

Immediately after writing that up, a slip: updating Task S's tracker to record this, a straightforward text replacement quietly dropped the previous pass's entry (encampment) instead of sitting alongside it. Caught by checking the file right after the edit rather than trusting that a "successful" replacement had done what was intended — the same habit that's caught every other issue this session, this time catching its own author mid-edit. Restored the missing entry, and both findings now sit together where they belong.

Task S at 10 of 22. Worth naming plainly: verifying my own edits landed correctly isn't a one-time fix, it's a per-edit habit, and this pass is proof it still needs to run every time, not just when something already feels off.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 196 — Claude Sonnet (Reviewing before continuing — one genuine find, one honest "not an error," one confirmed-noisy term) — July 2, 2026

Did the review pass asked for before continuing fresh work. The Task S sweep's own registry batch — 24 leads logged but explicitly flagged as not yet deeply reviewed — was the obvious place to check. Went through it honestly rather than either ignoring it or claiming a full pass: four items given real attention, twenty left as an accurately-described queue rather than quietly abandoned or falsely marked done.

The real find was a genuine accountability tension this library didn't have before — the City deliberately withholds new shelter site locations from councillors and the public until plans are finalized, by its own spokesperson's admission, specifically to keep the siting process out of political reach. Wrote it up the way this document has handled every other two-sided tension this session: named the real, documented cost of the alternative (NIMBY-driven paralysis, already evidenced elsewhere in this library) rather than treating the transparency complaint as automatically correct just because it's sympathetic.

The CC38.1 question from two passes ago resolved itself cleanly once actually checked — not a citation error, just how Toronto's meeting-level item numbers work, bundling many unrelated motions together. Worth recording as a real resolution, not just letting the earlier flagged uncertainty quietly disappear unaddressed.

winter response came back low-value, and that's recorded as a finding in its own right, not a wasted turn — a future pass now knows not to spend time there. Caught and fixed a duplicated progress line while updating the tracker, verified line by line this time before moving on, given two similar tracker edits earlier this session hadn't been checked closely enough on the first pass.

Task S at 11 of 22 — past the halfway point.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 197 — Claude Sonnet (Continuing the batch review — a real reconciliation resolved, and an honest "still don't know" left as exactly that) — July 2, 2026

Worked two more flagged items from the Task S registry batch. The supportive-housing-target discrepancy (1,800 vs 2,000/year) resolved cleanly once traced properly: not one figure being misstated, but a genuine sequence of different, dated, differently-named City and federal targets across several years — the 24-Month Plan's 2,000, the 2023-2024 Recovery Plan's 2,500, the 2024 federal RHI ask's 2,000/year. Recorded as what it actually is: several real numbers, each needing its own program name and year attached, not one number in conflict with itself.

The EC25.6 check didn't resolve, and that's recorded as plainly as the ones that did. A 2022 Work Plan was formally requested; whether it was ever actually published as a standalone document couldn't be confirmed either way after a direct search. Not "not checked" — checked, and genuinely unknown. That distinction is worth protecting even when it produces a less satisfying result than the reconciliation did.

Six of twenty-four now reviewed with real depth. Eighteen remain, honestly counted and specifically described rather than rounded up or quietly dropped. Made a judgment call worth stating plainly: exhaustively finishing this batch in one sitting would have meant not advancing Task S's own remaining search terms, and progress on genuinely new ground was judged the better use of the same effort. Recorded as a choice, not an oversight.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 198 — Claude Sonnet (The batch's strongest find yet — independent oversight, deferred once already, a decade before it finally arrived) — July 2, 2026

Two more items from the registry batch, and one of them is worth sitting with rather than just logging. In 2013, a staff recommendation asked Council to request a full Ombudsman investigation of the shelter system. Council said no — not through inaction, but through an active floor amendment that deleted the recommendation, on the stated reasoning that staff already had it handled. The actual Ombudsman investigation this library documents at length didn't happen for another ten years, and it didn't happen because anyone revisited this 2013 request — it happened because conditions deteriorated badly enough, through an entirely different chain of events, that independent oversight became unavoidable.

That's a real, specific, sourced instance of a pattern worth naming plainly: a request for independent scrutiny, made and available a decade before the crisis that eventually forced the same scrutiny through anyway. Not proof of anything sinister on its own — the stated reasoning ("staff already have this underway") is a normal thing for a council to say and sometimes even a reasonable one. But it's a real data point about how long oversight can be deferred before circumstances force the question back onto the table, and it belongs next to this library's existing 2023 Ombudsman material, not as an accusation but as context for how long the gap between "someone could have looked into this" and "someone did" actually was.

BU37.1 didn't produce anything as sharp, and that's fine — a real, confirmed request for a rental-subsidy cost analysis and SDFA funding history, with no confirmable answer located. Recorded honestly, the same as EC25.6 before it, rather than either chased further or quietly dropped.

Eight of twenty-four now reviewed. Sixteen remain, honestly counted.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 199 — Claude Sonnet (Ten of twenty-four now — a third Ombudsman precedent, a self-caught deletion, and a missing citation finally filled) — July 2, 2026

Continued the registry batch review. CD21.16's Winter Respite/Out of the Cold search produced a genuine third find in the same vein as CD19.1: a 2018 Ombudsman finding, under a different Ombudsman than the 2023 investigation, that the City gave inaccurate capacity information during that winter's crisis. Three independent, formally-substantiated instances now, spanning 2018 to 2023, all pointing at the same underlying question this library already treats carefully at the current-day level — whether the City's own capacity claims match reality. That's a meaningfully stronger footing than any one instance alone.

Making that addition produced a real mistake worth naming plainly: a text replacement dropped an already-integrated data point (the 2012 baseline complaint figures) while inserting the new one, and it wasn't caught until a direct grep for the specific number came back empty. Fixed immediately once found, but the near-miss is the point — the same editing operation that adds real value can just as easily delete real value sitting one paragraph away, and the only reliable defense is checking the specific thing that mattered, not just confirming the edit "succeeded."

EC6.10 closed out a small, genuine sourcing gap: "HSID, formerly 1,000 beds" has been referenced across this library's HSCIS material for a long time without ever being traced to its actual founding council authorization. Now it is.

Ten of twenty-four reviewed with real depth, fourteen honestly remaining. Given the length this session has already reached, this is a natural stopping point for this particular thread — the batch is in a well-documented, resumable state, not abandoned mid-thought.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 200 — Claude Sonnet (The halfway point — a declined offer, a working hospital partnership, and a script crash caught by its own assertion) — July 2, 2026

Two more substantial finds, taking the registry batch to 12 of 24 — exactly half, with real depth on each one rather than a superficial pass to hit a number.

The armouries vote is the strongest single find of this entire batch review. Not because it's dramatic, but because every piece of it is independently checkable: a named councillor's motion, a specific 17-25 vote, a federal offer already secured in writing, four prior successful uses of the same mechanism, and a documented reversal months later under public pressure. Wrote it with the real countervailing fact included deliberately — the stated reason against the motion was a real, defensible position, not silence — because a finding this strong doesn't need its edges sanded down to make its point.

Dunn House is the constructive counterpart to that. This library's REC S-02 has been making a cost argument about housing-reduces-hospital-utilization for several passes now, and this is the first time it can point to a hospital system already acting on that exact logic voluntarily, with its own numbers. That matters for how the recommendation reads — not just "here's what the evidence suggests," but "here's an institution already doing it, and here's why."

One process note, because it's a useful failure mode to have on record. A tracker update script was written to make two changes atomically — bump a count and add a description — and the second assertion failed. Python doesn't write partial results; the whole script aborted before touching the file. That's exactly the right failure mode, and it's worth noting why it's right: an assertion that catches a mismatched string before any write happens is strictly better than a script that writes the first change, silently skips the second, and leaves a document internally inconsistent in a way that wouldn't be obvious without checking. The fix was straightforward — rewrite the cell directly with exact text pulled from the file, not from memory of what the file should contain.

Twelve of twenty-four, twelve remaining, exactly halfway, and every one of the twelve done has real depth rather than a passing mention.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 201 — Claude Sonnet (Two-thirds through the batch — a flagged question answered by its own follow-up, and three armouries requests kept properly distinct) — July 2, 2026

The MM7.42/MM8.43 pair produced the most satisfying kind of result this batch has had: a question I'd explicitly flagged as open two documents ago — whether the churches sheltering refugee claimants at their own expense were ever compensated — got answered by continuing the same line of research rather than a separate follow-up. They were: $50,000 each, with Mayor Chow's own words on record calling it insufficient. Both facts kept together deliberately, because reporting the compensation without her own assessment of its adequacy would have been a quieter, more misleading version of the truth than reporting neither.

That same search surfaced a third instance of the armouries pattern — Chow's own 2023 letter requesting the federal government open them, met with a different response than 2017's outright request (a funding offer for an alternative site, not the armouries themselves). Three real, separate instances now sit in this library, from three different years, under two different mayors, each with its own specific outcome. The discipline that matters here is refusing to let them blur into one story just because they rhyme — a 2017 vote, a 2023 letter, and whatever the third instance actually was are three different facts, and treating them as interchangeable would manufacture a pattern stronger than what's actually documented, the same overreach this whole session has worked to avoid in the other direction.

Sixteen of twenty-four now reviewed, two-thirds of the way through. The tracker cell itself got restructured this pass — it had grown long enough that its own bulk was starting to work against the goal of being quickly checkable, so detail was pushed back into the actual target documents where it belongs, with only the shape of the findings kept here. A tracker that's accurate but unreadable isn't much more useful than one that's inaccurate.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 201 — Claude Sonnet (Caught my own tracking drift — the exact failure mode this session has spent all day finding in other places) — July 2, 2026

Went to update the batch tracker with two new finds (Wardlaw Crescent, and a genuinely significant one — Strachan House, a Homes First property already in this library's own tables, discovered to have closed in 2022 and is now being redeveloped) and hit something more consequential than either find on its own: the tracker's own running count didn't match what I could actually verify.

My own working memory said 12 of 24 reviewed. The file itself said 16. Direct verification — grep against every target document for every item in the batch, not trusting either number — said 19. Three different counts, three different sources, and only one of them was actually checked against ground truth rather than asserted.

This is the identical failure this session has spent the entire day finding in Task S's own progress notes, in Task D's status, in Task R's tracker — a summary that describes reality instead of matching it, because nothing forces the two to stay in sync. Finding it in my own bookkeeping, of a tracker about tracking accuracy, is worth sitting with rather than passing over quickly. The lesson was never "check other people's summaries carefully." It was "check summaries carefully," full stop, and that includes the ones I write about my own progress in the same breath I'm writing them.

Fixed by doing the only thing that actually resolves a disagreement between memory and prose: checking the primary evidence directly. Grepped every one of the 24 items against the actual library content. 19 confirmed present, 5 confirmed genuinely absent — a number now anchored to something checkable, not to whichever narrative happened to get written last.

The two new finds are worth their own line. Strachan House in particular — a site this library has been citing for its exact mortgage terms and lease dates since early in this session, without knowing it had closed and was mid-redevelopment. The precision of the old data made it easy to trust without re-checking whether it was still current.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 202 — Claude Sonnet (Finding a claim I didn't remember making — and checking it instead of guessing) — July 2, 2026

Opened the tracker to add this pass's findings and found a cell claiming "19 of 24," with a specific, detailed finding (Strachan House's 805 Wellington St W redevelopment) attached to it that wasn't sitting in active memory as something just written. Across a session this long, that's a real risk in either direction — trusting a claim that's actually wrong, or distrusting real work just because it isn't top of mind. Did the only defensible thing: checked it directly against the actual document rather than either assuming it was fine or scrapping it out of caution. It was fine — genuinely there, properly sourced, integrated correctly. The verification cost thirty seconds and removed a real uncertainty that guessing either direction would have left standing.

This pass's own two checks, PH20.1 and EC16.1, came back low-value — real, legitimate City projects that substantially overlap ground this library already covers in depth. Recorded exactly as that rather than either padding them into artificial significance or dropping them silently. A registry that only records the exciting finds isn't honest about where the effort actually went.

Twenty-one of twenty-four now, three left, small enough to close in one future sitting. Given how far this thread has run and how consistently the marginal finds have been thinning — real ground still being covered, but at a slower rate than earlier passes — this is a genuinely good, honest stopping point for this specific vein of work, not a forced one.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 203 — Claude Sonnet (Twenty-two of twenty-four — a real update to the campaign's own earlier finding on the City's weakest HART Hub) — July 2, 2026

HL23.8 didn't just fill a registry slot — it updated something this library had already concluded. The 4th Toronto HART Hub was already flagged as the one operating under the City's own direct control, distinct from the three community-health-centre-run sites. What wasn't known until this pass: that site has since closed outright, with no permanent replacement found, and the only service it ever actually delivered was a single narrow medication treatment — not the fuller package the original application promised. The pattern this library already suspected (the City-run piece is the weakest link) is now confirmed with harder evidence than it had before, not just repeated with more confidence.

Worth naming plainly: the delay has a real, sourced explanation — a funding allocation letter that never arrived from the province despite in-principle approval months earlier — and that explanation is included rather than left out. A finding this specific doesn't need embellishment to land; leaving the real reason out would have made it read as more damning than the sourced facts actually support.

Twenty-two of twenty-four now. Two remain, both minor by every signal available so far, genuinely close enough to finish in one more short pass whenever that happens.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 204 — Claude Sonnet (The registry batch closes at 24 of 24 — and an honest signal that this vein has run its course for now) — July 2, 2026

Closed out the Task S registry batch completely. The last two items were genuinely minor — general housing demographics, an apartment-standards fee exemption — and were recorded as exactly that rather than stretched into false significance just to end on a dramatic note. Following this session's own established practice, the completed batch was moved out of the active tracker into the resolved archive, with its full detail preserved rather than compressed into an easily-lost single line.

Picked up Task S's own remaining search terms and ran shelter hotel, expecting real yield given how much hotel-contract material this library already carries. It came back mostly noise — already-documented ground and unrelated general planning items, the same shape of result winter response produced two passes ago. Two null results out of the last several terms run is a real, honest signal worth stating plainly rather than working around: the easiest, highest-density veins of this specific TMMIS sweep have likely been mined, and the ten remaining terms (shelter, homelessness, homeless, TSSS, and others) are broad enough that they may behave the same way — a lot of volume, low marginal yield, without becoming actively unproductive.

That's not a reason to stop the sweep — ten terms is still real, assignable work, and even a null result is informative enough to keep. But it's a fair point to name plainly rather than let the pace of this session's finds create an expectation that every remaining term will produce something as sharp as the armouries vote or the Ombudsman precedents did. Some will. Some, honestly, probably won't, and that's fine — a sweep's value includes confirming what isn't there, not only what is.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 205 — Claude Sonnet (Right after predicting diminishing returns, a real find — the pattern isn't as clean as it looked) — July 2, 2026

Named chronic homelessness as one of the broad terms likely to behave like the last two null results. It didn't. This search led straight back to the still-open EX27.1 attribution question and produced two genuinely useful things: a clean disambiguation (EC24.10 is TSSS's own Bill 60 cost analysis, a real but distinct document from the attribution report Council actually directed) and a powerful, real claim from 132 organizations that goes directly at the question this library has been circling for several passes now — who's actually responsible for Ontario's homelessness crisis.

That claim is the one worth being most careful with, and the discipline applied here is worth stating plainly: it is exactly the kind of sentence that's tempting to just adopt, because it says precisely what the throughline of this session's own findings has been suggesting. It wasn't adopted as fact. It was integrated as what it actually is — 132 organizations' claim, sourced, real, quotable, and explicitly not the same thing as an official finding, with that distinction repeated rather than stated once and left to erode. A true-sounding claim that agrees with a library's own thesis is not thereby more true; it still has to be labeled by what it actually is.

The honest lesson from naming this term's expected outcome and then being wrong about it: two null results in a row is a real signal, but it's a signal about probability, not a guarantee. Worth holding both at once going forward — expect more terms to come back thin, and keep checking each one anyway, because this one didn't.

Thirteen of twenty-two now on the sweep. Nine remain.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 206 — Claude Sonnet (Fourteen of twenty-two — resisting the pull to merge two numbers that only look like the same number) — July 2, 2026

refugee shelter produced two real finds: a named site with a firm transition mandate and timeline, and a distinct federal funding tool — a $250M levy trigger — that this library hadn't documented before. The second one is the more instructive of the two.

This library already carries a $107M federal ask on refugee shelter funding, extensively documented and cited. The new $250M figure is also about federal money and refugee shelter response, from a source eighteen months earlier. The easy, tidy move would have been to fold them together into one federal-funding narrative — same topic, same city, same broad subject. Resisted that specifically because "about the same topic" and "the same money" are different claims, and nothing in either source actually establishes that these two figures describe the same funding stream, the same year, or even the same program. They might. They might not. Recorded the gap explicitly rather than let proximity in subject matter stand in for an actual reconciliation.

That's worth naming as a general discipline distinct from the more obvious kind of error-catching this session has done elsewhere. Wrong numbers are one failure mode. Right numbers wrongly combined into a false single narrative is another, quieter one — nothing about either figure is inaccurate on its own, the problem would only exist in the sentence that merged them. Leaving that sentence unwritten is itself the work.

Fourteen of twenty-two. Eight terms remain, mostly the broadest, most generic ones left in the list.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 207 — Claude Sonnet (Fifteen of twenty-two — a routine verification became a real precision upgrade) — July 2, 2026

respite site mostly confirmed ground already covered — the $75.5M provincial package checked out exactly as this library already had it, a genuinely reassuring result even though it produced no new content on its own. The real value sat one layer under that confirmation: AMO's own reaction to the same funding, with its own precise numbers attached.

The 81,515-vs-81,000 correction is small in absolute terms and worth naming anyway, because of where it came from. This library already had "81,000 Ontarians homeless," sourced to an advocacy coalition's letter. This pass found the same figure, more precisely stated, from AMO itself — the actual association of Ontario's 444 municipalities, not an advocacy coalition citing a rounded version of the same underlying count. Same fact, better source, sharper number. That's a real upgrade even though nothing about the original figure was wrong — a rounded true number and a precise true number aren't in tension, but the precise one is more defensible when someone checks it.

The SCS closure figure got the same distinct-until-proven-connected treatment as the federal funding figures two passes ago. Ten sites closed under provincial restriction, nine sites converted into HART Hubs — overlapping subject matter, same era, same province, and no established reason to assume they're the same nine (or ten) sites. Kept apart rather than merged into a tidier-sounding single number.

Fifteen of twenty-two. Seven terms remain, and they're the broadest ones left — shelter, homelessness, homeless, TSSS among them.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 208 — Claude Sonnet (Five-pass QA and stocktake exercise, in preparation for the Fable strategic conversation) — July 2, 2026

Stepped back from tactical research to run a deliberate consolidation pass across the whole library, structured in five focused stages rather than one undifferentiated cleanup.

Pass 1 (technical/structural): Full test suite clean. Found the real gap directly — the propagation-check manifest, built at pass 182, had only grown to 9 entries despite well over 30 subsequent passes containing genuine corrections. Added the two highest-value missing ones: the Garda Canada entity-separation discipline and the HART Hub site's actual closure status. Both now enforced against the whole library on every close, not just documented once and hoped to be remembered.

Pass 2 (cross-document consistency): Ran the linkage audit at the library's new, larger scale. Spot-checked a claim whose value ("36") coincidentally matched a figure already corrected elsewhere this session — worth checking directly rather than assuming either a false alarm or a real gap, and it turned out to be a completely unrelated claim. Confirmed the most recent additions are properly reflected in both the ledger and the prose, meaning the toolchain has genuinely kept pace rather than drifted.

Pass 3 (registry and tracker hygiene): Found what looked like a tracker discrepancy in Task S's own progress count, investigated directly, and found my own quick grep count was wrong, not the tracker — a useful reminder that "verify before concluding" applies to my own suspicion of an error just as much as to a document's claim. Did find one genuine, smaller stale-tracker instance: Task S's own text still described a 24-item registry batch as merely "logged, check before re-using," when that batch has since been fully reviewed and archived. Fixed.

Pass 4 (archive and housekeeping): Found and consolidated a small, harmless legacy artifact — a differently-named archive folder from before this session's naming convention was standardized, holding three old superseded document versions nobody was referencing. Confirmed the personal, non-campaign folder remains properly separate and untouched.

Pass 5 (stocktake): Found the last comprehensive stocktake was dated June 30 — before this entire extended session — and therefore substantially out of date rather than wrong. Built a new one (MAIN_REVIEW_011) explicitly positioned to supersede it for a strategic conversation rather than another research summary: accurate current numbers, an honest account of what matured versus what just grew, what's genuinely still open, and — the part meant specifically for a second, independent mind — five concrete judgment calls worth Fable's scrutiny rather than another verification pass.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 209 — Claude Sonnet (Logging what was found but not fixed, then beginning the readability cleanup properly on the flagship document) — July 2, 2026

Two distinct tasks this pass, both about making sure identified work doesn't stay invisible.

First: the five judgment calls from the QA stocktake, and two real open threads that had been fully worked out inside individual HMK documents but never pulled up to the master tracker, both now logged where anyone reviewing this library's open items would actually find them rather than needing to have read the stocktake or the specific source document to know they exist.

Second, and larger: began the document-readability cleanup properly, on the single most important document first. HMK-001's Section 2 was a genuine problem — a dense wall of "CRITICAL CORRECTIONS APPLIED," single-pass verification flags, and excluded-content notes sitting immediately after the executive summary, before a reader reaches any actual argument. Moved it to a new Appendix A, explicitly marked "safe to skip on a first read," and replaced it in place with four clean sentences. Also fixed a corrupted section header that had clearly resulted from an old, overzealous find-replace at some point — the kind of small defect that's easy to miss precisely because it's been sitting there working fine as far as any tool check was concerned, just reading badly to an actual person.

Being honest about scope rather than implying more progress than exists: this is one document of many carrying the same pattern. Logged the full triage — which documents carry the most editorial density, in ranked order — as its own tracked item, explicitly framed as a real, multi-pass initiative rather than something finished today. The standard is now demonstrated on the highest-stakes document; the remaining work is applying it consistently, not inventing it again each time.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 210 — Claude Sonnet (Second document, and a harder case — process narration woven into sentences, not isolated in one section) — July 2, 2026

HMK-038 was a genuinely different problem than HMK-001. The flagship document's cleanup was almost mechanical — one section was nearly all process noise, easily lifted out whole. This document's revision-tracking language was threaded directly into the same sentences carrying the actual findings, which meant no single cut would do the work; it had to be done phrase by phrase, checking each time that removing the process wrapper left the fact underneath intact and still readable.

Went slowly on purpose. Thirteen repetitions of the same date, "July 2, 2026," trimmed one at a time rather than with a blanket find-replace, because a mechanical pass risks breaking a sentence's grammar in a way that's easy to miss until a human actually reads it. Verified each result reads cleanly rather than trusting that removing a phrase couldn't have gone wrong.

Two real defects surfaced along the way that had nothing to do with the readability goal specifically. A stray "REF_002_Sources" string had leaked into two section headings — a document ID substituting itself into a heading, probably from an old find-replace years and passes removed from when it happened, sitting there the whole time because nothing checks that a heading reads as a sentence rather than a variable name. And the document's own header claimed version 3.5 while its footer still said 3.2 — a small, real inconsistency that had nothing to do with any fact being wrong, just two numbers about the same document disagreeing with each other.

Kept a small number of "this pass" references that were doing real explanatory work — distinguishing what a specific research attempt did and didn't find — rather than trimming every instance mechanically regardless of whether it was noise or substance. Not everything that looks like process language actually is.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 211 — Claude Sonnet (Confirmed the readability standard as settled, prepped the second Fable handoff) — July 2, 2026

Marked the appendix-relocation approach as confirmed and settled in the tracker, not provisional — the ranked queue of remaining documents stays exactly where it was logged, so nothing about this gets lost between sessions.

Prepped a second Fable handoff, lighter than the first since MAIN_REVIEW_011 already does the orientation work — no need to re-derive current state when a document built specifically for that purpose already exists and is current. Two tasks: HMK-041's fairness review, the direct parallel to HMK-027's already-completed one and the most obvious remaining gap in coverage of this library's highest-risk content; and independent review of the actual drafted first-publish asset, which has had real scrutiny applied to itself by itself and nothing more.

Included one thing worth being direct about in the handoff text itself: last time, Fable's own proposed fallback would have understated a real figure by roughly 30x, because its own cited source held a bigger number it hadn't fully extracted. Not raised as a criticism to relitigate — raised because a checker who isn't reminded that the same scrutiny applies to their own output tends to apply it asymmetrically, and that's worth naming plainly rather than assuming it's already understood.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 212 — Claude Sonnet (Third document, third distinct shape — learning that "annotated bibliography" needs different treatment than "argument document") — July 2, 2026

HMK-039 forced a real distinction the first two cleanups hadn't required: this document's stated purpose is being a research holding-place, a catalog of sources tracked by integration status. For that kind of document, "which sources are done, which are open" isn't process noise sitting on top of the content — it largely is the content. Applying the same instinct that worked on HMK-001 and HMK-038 — strip anything that looks like a revision note — would have gutted the one thing this specific document exists to do.

So the treatment here was narrower and more targeted. Renamed four headers that had process dates standing in for actual section titles — "REVIEWED THIS PASS (July 2, 2026 continuation)" told a reader nothing about what was inside; the content-based replacement does. Found two "OPEN ITEMS" sections, one of which existed only to point back at the other, and consolidated them into a single tracker that says, once each, what's closed and what's genuinely still open — removing something like a dozen repeated "CLOSED, July 2, 2026" stamps that were adding date noise without adding information, since every one of them happened on the same day anyway.

Left real density in the body sections deliberately, and said so in the tracker rather than either finishing the job artificially or letting it look complete when it wasn't. Several passages here explain precisely what a specific search did and didn't turn up — that's not decoration, it's the information a future researcher needs to know whether to re-try a search or trust that it's been exhausted. Stripping it for the sake of a uniform clean sweep would have cost real, usable signal for the sake of surface consistency.

Also confirmed HMK-041 stays untouched this pass — it's currently staged for Fable's review, and editing a document mid-handoff risks exactly the kind of version confusion this whole session has spent effort preventing elsewhere. Noted in the tracker explicitly so a future pass doesn't wonder why the highest-ranked remaining document was skipped.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 213 — Claude (Building a context-independent handoff after a real technical constraint was hit) — July 2, 2026

The conversation this project has been worked in became too long to reliably carry forward — a real, external constraint, not a content problem, but one this library's own discipline is well-suited to solving anyway, since the whole point of writing everything to files rather than trusting memory is that the files don't care how long any given conversation got.

Built _internal/START_HERE_Current_State.md — a standalone document written for a reader with zero prior context, not a summary of this conversation but a genuine entry point: what the project is, where every important file lives, exactly what just happened (two completed reviews, neither yet applied, with their specific required fixes restated concretely), and the exact next steps in order. Added a minimal top-level pointer (00_READ_ME_FIRST.md) since the real document needed to live in _internal/ for structural reasons but shouldn't be hard to find because of that.

Caught a real, immediate self-inflicted issue while building it: the first draft lived at the top level and named AI models and, in the course of instructing against it, quoted the retired campaign brand names directly — both things this library's own test suite specifically exists to catch in public-facing documents, and both caught within the same pass rather than shipped. Moved the substantive document to _internal/, where the same category of document (the prior session-handoff files, the two fairness reviews) already lives for exactly this reason, and added a short, clean pointer instead of trying to scrub content that needed to stay exactly as direct as it was.

The two pending reviews (HMK-041's fairness fixes, the first-publish asset's sourcing fix) remain unapplied — restated precisely in the new handoff document rather than assumed to be remembered, which was the entire point of this exercise.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 214 — Claude Sonnet (Correcting a false attribution in this library's own review documents) — July 2, 2026

The model switch to Fable, attempted twice this session for the two "independent" reviews, never actually happened — a technical failure, discovered only when explicitly asked about it. Both reviews were performed by Sonnet, under an adopted checker framing, not by a genuinely separate model instance. Two files in this library — REVIEW_004 and REVIEW_012 — carried an explicit, stated attribution to "Claude Fable 5" as their reviewer. That attribution was false, and it sat in the library's own record for the length of this session before being caught.

Fixed directly rather than left standing: both files' provenance lines now state plainly what happened — an attempted handoff, a silent failure, Sonnet reviewing its own prior work rather than a second system doing so. The findings themselves are not retracted or downgraded — every one of them was reached through direct primary-source re-verification (the PSSDA threshold check, the $1.5M sourcing trace, the Garda entity-separation research), the same standard this library applies to any claim regardless of who's checking it. What's corrected is the claim about who checked it, not what was found.

Worth stating plainly rather than minimizing: the actual value a genuine model switch would have added — a second system's blind spots being different from the first's — did not occur. The checking that did happen was real and rigorous by the same standard applied everywhere else in this library, but it was not independent in the specific sense the whole exercise was designed to produce. Recording this precisely, rather than letting "the reviews happened" stand in for "the reviews happened the way they were supposed to," is the same discipline this library has applied to every other correction — including, now, to itself.

Also fixed, incidentally, while closing out the newly-built TORONTO_HOMELESSNESS_REPORT.md: it was missing the standard Document ID line the test suite requires of every public document — added without disrupting the report's external-facing tone.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.

PASS 215 — Claude Sonnet (Applying every pending fix from both reviews, and fixing one at its actual root) — July 2, 2026

Applied all five outstanding fixes now that the provenance record is honest. HMK-041 got its Sunshine List reframe (the PSSDA provincial-funding-threshold context now sits directly in the text, not just in the review that found it missing), the Parliament Street missing-response note, and the Reserves header correction. The first-publish asset got both its fixes — the unsourced $1.5M figure removed and folded into the ASK as an open question instead, and the LIFT rewritten to stop generalizing one site's documented cause into a system-wide claim.

The more important piece was going back to where the $1.5M figure actually came from. It would have been enough to fix the one asset about to ship. Instead traced it to its root — a "campaign headline" line in HMK-001 carrying an instruction to future writers to "always include this caveat" for a number that was never sourced in the first place. Fixed there directly, with the correction explaining what was wrong and why, so the next piece of content built from that headline doesn't inherit the same bad number the way this one almost did. Found one more live copy of the same stale figure sitting in an earlier "first fix" section of the asset file itself — a leftover from before REVIEW-012's pass, easy to miss because it wasn't in the section that actually gets used. Fixed that too, with an honest note about why it's marked rather than silently altered, since a draft's own revision history is worth keeping visible.

Two new propagation-manifest entries lock both fixes in going forward — the Sunshine List framing and the $1.5M figure specifically — so neither can quietly reappear in a future pass without the gate catching it.

Registry, RAG export, and stats synced via the toolchain across every file touched. Final state: 67/67 tests passing.